diff --git a/src/xxHash/xxh3.h b/src/xxHash/xxh3.h new file mode 100644 index 00000000..b3a82958 --- /dev/null +++ b/src/xxHash/xxh3.h @@ -0,0 +1,1613 @@ +/* + xxHash - Extremely Fast Hash algorithm + Development source file for `xxh3` + Copyright (C) 2019-present, Yann Collet. + + BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php) + + Redistribution and use in source and binary forms, with or without + modification, are permitted provided that the following conditions are + met: + + * Redistributions of source code must retain the above copyright + notice, this list of conditions and the following disclaimer. + * Redistributions in binary form must reproduce the above + copyright notice, this list of conditions and the following disclaimer + in the documentation and/or other materials provided with the + distribution. + + THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS + "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT + LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR + A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT + OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, + SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT + LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE + OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + + You can contact the author at : + - xxHash source repository : https://github.com/Cyan4973/xxHash +*/ + +/* Note : + This file is separated for development purposes. + It will be integrated into `xxhash.c` when development phase is complete. +*/ + +#ifndef XXH3_H +#define XXH3_H + + +/* === Dependencies === */ + +#undef XXH_INLINE_ALL /* in case it's already defined */ +#define XXH_INLINE_ALL +#include "xxhash.h" + + +/* === Compiler specifics === */ + +#if defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L /* >= C99 */ +# define XXH_RESTRICT restrict +#else +/* note : it might be useful to define __restrict or __restrict__ for some C++ compilers */ +# define XXH_RESTRICT /* disable */ +#endif + +#if defined(__GNUC__) +# if defined(__AVX2__) +# include +# elif defined(__SSE2__) +# include +# elif defined(__ARM_NEON__) || defined(__ARM_NEON) +# define inline __inline__ /* clang bug */ +# include +# undef inline +# endif +#elif defined(_MSC_VER) +# include +#endif + +/* + * Sanity check. + * + * XXH3 only requires these features to be efficient: + * + * - Usable unaligned access + * - A 32-bit or 64-bit ALU + * - If 32-bit, a decent ADC instruction + * - A 32 or 64-bit multiply with a 64-bit result + * + * Almost all 32-bit and 64-bit targets meet this, except for Thumb-1, the + * classic 16-bit only subset of ARM's instruction set. + * + * First of all, Thumb-1 lacks support for the UMULL instruction which + * performs the important long multiply. This means numerous __aeabi_lmul + * calls. + * + * Second of all, the 8 functional registers are just not enough. + * Setup for __aeabi_lmul, byteshift loads, pointers, and all arithmetic need + * Lo registers, and this shuffling results in thousands more MOVs than A32. + * + * A32 and T32 don't have this limitation. They can access all 14 registers, + * do a 32->64 multiply with UMULL, and the flexible operand is helpful too. + * + * If compiling Thumb-1 for a target which supports ARM instructions, we + * will give a warning. + * + * Usually, if this happens, it is because of an accident and you probably + * need to specify -march, as you probably meant to compileh for a newer + * architecture. + */ +#if defined(__thumb__) && !defined(__thumb2__) && defined(__ARM_ARCH_ISA_ARM) +# warning "XXH3 is highly inefficient without ARM or Thumb-2." +#endif + +/* ========================================== + * Vectorization detection + * ========================================== */ +#define XXH_SCALAR 0 +#define XXH_SSE2 1 +#define XXH_AVX2 2 +#define XXH_NEON 3 +#define XXH_VSX 4 + +#ifndef XXH_VECTOR /* can be defined on command line */ +# if defined(__AVX2__) +# define XXH_VECTOR XXH_AVX2 +# elif defined(__SSE2__) || defined(_M_AMD64) || defined(_M_X64) || (defined(_M_IX86_FP) && (_M_IX86_FP == 2)) +# define XXH_VECTOR XXH_SSE2 +# elif defined(__GNUC__) /* msvc support maybe later */ \ + && (defined(__ARM_NEON__) || defined(__ARM_NEON)) \ + && (defined(__LITTLE_ENDIAN__) /* We only support little endian NEON */ \ + || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)) +# define XXH_VECTOR XXH_NEON +# elif defined(__PPC64__) && defined(__POWER8_VECTOR__) && defined(__GNUC__) +# define XXH_VECTOR XXH_VSX +# else +# define XXH_VECTOR XXH_SCALAR +# endif +#endif + +/* control alignment of accumulator, + * for compatibility with fast vector loads */ +#ifndef XXH_ACC_ALIGN +# if XXH_VECTOR == 0 /* scalar */ +# define XXH_ACC_ALIGN 8 +# elif XXH_VECTOR == 1 /* sse2 */ +# define XXH_ACC_ALIGN 16 +# elif XXH_VECTOR == 2 /* avx2 */ +# define XXH_ACC_ALIGN 32 +# elif XXH_VECTOR == 3 /* neon */ +# define XXH_ACC_ALIGN 16 +# elif XXH_VECTOR == 4 /* vsx */ +# define XXH_ACC_ALIGN 16 +# endif +#endif + +/* xxh_u64 XXH_mult32to64(xxh_u32 a, xxh_u64 b) { return (xxh_u64)a * (xxh_u64)b; } */ +#if defined(_MSC_VER) && defined(_M_IX86) +# include +# define XXH_mult32to64(x, y) __emulu(x, y) +#else +# define XXH_mult32to64(x, y) ((xxh_u64)((x) & 0xFFFFFFFF) * (xxh_u64)((y) & 0xFFFFFFFF)) +#endif + +/* VSX stuff. It's a lot because VSX support is mediocre across compilers and + * there is a lot of mischief with endianness. */ +#if XXH_VECTOR == XXH_VSX +# include +# undef vector +typedef __vector unsigned long long U64x2; +typedef __vector unsigned char U8x16; +typedef __vector unsigned U32x4; + +#ifndef XXH_VSX_BE +# if defined(__BIG_ENDIAN__) \ + || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__) +# define XXH_VSX_BE 1 +# elif defined(__VEC_ELEMENT_REG_ORDER__) && __VEC_ELEMENT_REG_ORDER__ == __ORDER_BIG_ENDIAN__ +# warning "-maltivec=be is not recommended. Please use native endianness." +# define XXH_VSX_BE 1 +# else +# define XXH_VSX_BE 0 +# endif +#endif + +/* We need some helpers for big endian mode. */ +#if XXH_VSX_BE +/* A wrapper for POWER9's vec_revb. */ +# ifdef __POWER9_VECTOR__ +# define XXH_vec_revb vec_revb +# else +XXH_FORCE_INLINE U64x2 XXH_vec_revb(U64x2 val) +{ + U8x16 const vByteSwap = { 0x07, 0x06, 0x05, 0x04, 0x03, 0x02, 0x01, 0x00, + 0x0F, 0x0E, 0x0D, 0x0C, 0x0B, 0x0A, 0x09, 0x08 }; + return vec_perm(val, val, vByteSwap); +} +# endif + +/* Power8 Crypto gives us vpermxor which is very handy for + * PPC64EB. + * + * U8x16 vpermxor(U8x16 a, U8x16 b, U8x16 mask) + * { + * U8x16 ret; + * for (int i = 0; i < 16; i++) { + * ret[i] = a[mask[i] & 0xF] ^ b[mask[i] >> 4]; + * } + * return ret; + * } + * + * Because both of the main loops load the key, swap, and xor it with input, + * we can combine the key swap into this instruction. + */ +# ifdef vec_permxor +# define XXH_vec_permxor vec_permxor +# else +# define XXH_vec_permxor __builtin_crypto_vpermxor +# endif +#endif +/* + * Because we reinterpret the multiply, there are endian memes: vec_mulo actually becomes + * vec_mule. + * + * Additionally, the intrinsic wasn't added until GCC 8, despite existing for a while. + * Clang has an easy way to control this, we can just use the builtin which doesn't swap. + * GCC needs inline assembly. */ +#if __has_builtin(__builtin_altivec_vmuleuw) +# define XXH_vec_mulo __builtin_altivec_vmulouw +# define XXH_vec_mule __builtin_altivec_vmuleuw +#else +/* Adapted from https://github.com/google/highwayhash/blob/master/highwayhash/hh_vsx.h. */ +XXH_FORCE_INLINE U64x2 XXH_vec_mulo(U32x4 a, U32x4 b) { + U64x2 result; + __asm__("vmulouw %0, %1, %2" : "=v" (result) : "v" (a), "v" (b)); + return result; +} +XXH_FORCE_INLINE U64x2 XXH_vec_mule(U32x4 a, U32x4 b) { + U64x2 result; + __asm__("vmuleuw %0, %1, %2" : "=v" (result) : "v" (a), "v" (b)); + return result; +} +#endif +#endif + + +/* ========================================== + * XXH3 default settings + * ========================================== */ + +#define XXH_SECRET_DEFAULT_SIZE 192 /* minimum XXH3_SECRET_SIZE_MIN */ + +#if (XXH_SECRET_DEFAULT_SIZE < XXH3_SECRET_SIZE_MIN) +# error "default keyset is not large enough" +#endif + +XXH_ALIGN(64) static const xxh_u8 kSecret[XXH_SECRET_DEFAULT_SIZE] = { + 0xb8, 0xfe, 0x6c, 0x39, 0x23, 0xa4, 0x4b, 0xbe, 0x7c, 0x01, 0x81, 0x2c, 0xf7, 0x21, 0xad, 0x1c, + 0xde, 0xd4, 0x6d, 0xe9, 0x83, 0x90, 0x97, 0xdb, 0x72, 0x40, 0xa4, 0xa4, 0xb7, 0xb3, 0x67, 0x1f, + 0xcb, 0x79, 0xe6, 0x4e, 0xcc, 0xc0, 0xe5, 0x78, 0x82, 0x5a, 0xd0, 0x7d, 0xcc, 0xff, 0x72, 0x21, + 0xb8, 0x08, 0x46, 0x74, 0xf7, 0x43, 0x24, 0x8e, 0xe0, 0x35, 0x90, 0xe6, 0x81, 0x3a, 0x26, 0x4c, + 0x3c, 0x28, 0x52, 0xbb, 0x91, 0xc3, 0x00, 0xcb, 0x88, 0xd0, 0x65, 0x8b, 0x1b, 0x53, 0x2e, 0xa3, + 0x71, 0x64, 0x48, 0x97, 0xa2, 0x0d, 0xf9, 0x4e, 0x38, 0x19, 0xef, 0x46, 0xa9, 0xde, 0xac, 0xd8, + 0xa8, 0xfa, 0x76, 0x3f, 0xe3, 0x9c, 0x34, 0x3f, 0xf9, 0xdc, 0xbb, 0xc7, 0xc7, 0x0b, 0x4f, 0x1d, + 0x8a, 0x51, 0xe0, 0x4b, 0xcd, 0xb4, 0x59, 0x31, 0xc8, 0x9f, 0x7e, 0xc9, 0xd9, 0x78, 0x73, 0x64, + + 0xea, 0xc5, 0xac, 0x83, 0x34, 0xd3, 0xeb, 0xc3, 0xc5, 0x81, 0xa0, 0xff, 0xfa, 0x13, 0x63, 0xeb, + 0x17, 0x0d, 0xdd, 0x51, 0xb7, 0xf0, 0xda, 0x49, 0xd3, 0x16, 0x55, 0x26, 0x29, 0xd4, 0x68, 0x9e, + 0x2b, 0x16, 0xbe, 0x58, 0x7d, 0x47, 0xa1, 0xfc, 0x8f, 0xf8, 0xb8, 0xd1, 0x7a, 0xd0, 0x31, 0xce, + 0x45, 0xcb, 0x3a, 0x8f, 0x95, 0x16, 0x04, 0x28, 0xaf, 0xd7, 0xfb, 0xca, 0xbb, 0x4b, 0x40, 0x7e, +}; + +/* + * GCC for x86 has a tendency to use SSE in this loop. While it + * successfully avoids swapping (as MUL overwrites EAX and EDX), it + * slows it down because instead of free register swap shifts, it + * must use pshufd and punpckl/hd. + * + * To prevent this, we use this attribute to shut off SSE. + */ +#if defined(__GNUC__) && !defined(__clang__) && defined(__i386__) +__attribute__((__target__("no-sse"))) +#endif +static XXH128_hash_t +XXH_mult64to128(xxh_u64 lhs, xxh_u64 rhs) +{ + /* + * GCC/Clang __uint128_t method. + * + * On most 64-bit targets, GCC and Clang define a __uint128_t type. + * This is usually the best way as it usually uses a native long 64-bit + * multiply, such as MULQ on x86_64 or MUL + UMULH on aarch64. + * + * Usually. + * + * Despite being a 32-bit platform, Clang (and emscripten) define this + * type despite not having the arithmetic for it. This results in a + * laggy compiler builtin call which calculates a full 128-bit multiply. + * In that case it is best to use the portable one. + * https://github.com/Cyan4973/xxHash/issues/211#issuecomment-515575677 + */ +#if defined(__GNUC__) && !defined(__wasm__) \ + && defined(__SIZEOF_INT128__) \ + || (defined(_INTEGRAL_MAX_BITS) && _INTEGRAL_MAX_BITS >= 128) + + __uint128_t product = (__uint128_t)lhs * (__uint128_t)rhs; + XXH128_hash_t const r128 = { (xxh_u64)(product), (xxh_u64)(product >> 64) }; + return r128; + + /* + * MSVC for x64's _umul128 method. + * + * xxh_u64 _umul128(xxh_u64 Multiplier, xxh_u64 Multiplicand, xxh_u64 *HighProduct); + * + * This compiles to single operand MUL on x64. + */ +#elif defined(_M_X64) || defined(_M_IA64) + +#ifndef _MSC_VER +# pragma intrinsic(_umul128) +#endif + xxh_u64 product_high; + xxh_u64 const product_low = _umul128(lhs, rhs, &product_high); + XXH128_hash_t const r128 = { product_low, product_high }; + return r128; + +#else + /* + * Portable scalar method. Optimized for 32-bit and 64-bit ALUs. + * + * This is a fast and simple grade school multiply, which is shown + * below with base 10 arithmetic instead of base 0x100000000. + * + * 9 3 // D2 lhs = 93 + * x 7 5 // D2 rhs = 75 + * ---------- + * 1 5 // D2 lo_lo = (93 % 10) * (75 % 10) + * 4 5 | // D2 hi_lo = (93 / 10) * (75 % 10) + * 2 1 | // D2 lo_hi = (93 % 10) * (75 / 10) + * + 6 3 | | // D2 hi_hi = (93 / 10) * (75 / 10) + * --------- + * 2 7 | // D2 cross = (15 / 10) + (45 % 10) + 21 + * + 6 7 | | // D2 upper = (27 / 10) + (45 / 10) + 63 + * --------- + * 6 9 7 5 + * + * The reasons for adding the products like this are: + * 1. It avoids manual carry tracking. Just like how + * (9 * 9) + 9 + 9 = 99, the same applies with this for + * UINT64_MAX. This avoids a lot of complexity. + * + * 2. It hints for, and on Clang, compiles to, the powerful UMAAL + * instruction available in ARMv6+ A32/T32, which is shown below: + * + * void UMAAL(xxh_u32 *RdLo, xxh_u32 *RdHi, xxh_u32 Rn, xxh_u32 Rm) + * { + * xxh_u64 product = (xxh_u64)*RdLo * (xxh_u64)*RdHi + Rn + Rm; + * *RdLo = (xxh_u32)(product & 0xFFFFFFFF); + * *RdHi = (xxh_u32)(product >> 32); + * } + * + * This instruction was designed for efficient long multiplication, + * and allows this to be calculated in only 4 instructions which + * is comparable to some 64-bit ALUs. + * + * 3. It isn't terrible on other platforms. Usually this will be + * a couple of 32-bit ADD/ADCs. + */ + + /* First calculate all of the cross products. */ + xxh_u64 const lo_lo = XXH_mult32to64(lhs & 0xFFFFFFFF, rhs & 0xFFFFFFFF); + xxh_u64 const hi_lo = XXH_mult32to64(lhs >> 32, rhs & 0xFFFFFFFF); + xxh_u64 const lo_hi = XXH_mult32to64(lhs & 0xFFFFFFFF, rhs >> 32); + xxh_u64 const hi_hi = XXH_mult32to64(lhs >> 32, rhs >> 32); + + /* Now add the products together. These will never overflow. */ + xxh_u64 const cross = (lo_lo >> 32) + (hi_lo & 0xFFFFFFFF) + lo_hi; + xxh_u64 const upper = (hi_lo >> 32) + (cross >> 32) + hi_hi; + xxh_u64 const lower = (cross << 32) | (lo_lo & 0xFFFFFFFF); + + XXH128_hash_t r128 = { lower, upper }; + return r128; +#endif +} + +/* + * We want to keep the attribute here because a target switch + * disables inlining. + * + * Does a 64-bit to 128-bit multiply, then XOR folds it. + * The reason for the separate function is to prevent passing + * too many structs around by value. This will hopefully inline + * the multiply, but we don't force it. + */ +#if defined(__GNUC__) && !defined(__clang__) && defined(__i386__) +__attribute__((__target__("no-sse"))) +#endif +static xxh_u64 +XXH3_mul128_fold64(xxh_u64 lhs, xxh_u64 rhs) +{ + XXH128_hash_t product = XXH_mult64to128(lhs, rhs); + return product.low64 ^ product.high64; +} + + +static XXH64_hash_t XXH3_avalanche(xxh_u64 h64) +{ + h64 ^= h64 >> 37; + h64 *= PRIME64_3; + h64 ^= h64 >> 32; + return h64; +} + + +/* ========================================== + * Short keys + * ========================================== */ + +XXH_FORCE_INLINE XXH64_hash_t +XXH3_len_1to3_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed) +{ + XXH_ASSERT(input != NULL); + XXH_ASSERT(1 <= len && len <= 3); + XXH_ASSERT(secret != NULL); + { xxh_u8 const c1 = input[0]; + xxh_u8 const c2 = input[len >> 1]; + xxh_u8 const c3 = input[len - 1]; + xxh_u32 const combined = ((xxh_u32)c1) | (((xxh_u32)c2) << 8) | (((xxh_u32)c3) << 16) | (((xxh_u32)len) << 24); + xxh_u64 const keyed = (xxh_u64)combined ^ (XXH_readLE32(secret) + seed); + xxh_u64 const mixed = keyed * PRIME64_1; + return XXH3_avalanche(mixed); + } +} + +XXH_FORCE_INLINE XXH64_hash_t +XXH3_len_4to8_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed) +{ + XXH_ASSERT(input != NULL); + XXH_ASSERT(secret != NULL); + XXH_ASSERT(4 <= len && len <= 8); + { xxh_u32 const input_lo = XXH_readLE32(input); + xxh_u32 const input_hi = XXH_readLE32(input + len - 4); + xxh_u64 const input_64 = input_lo | ((xxh_u64)input_hi << 32); + xxh_u64 const keyed = input_64 ^ (XXH_readLE64(secret) + seed); + xxh_u64 const mix64 = len + ((keyed ^ (keyed >> 51)) * PRIME32_1); + return XXH3_avalanche((mix64 ^ (mix64 >> 47)) * PRIME64_2); + } +} + +XXH_FORCE_INLINE XXH64_hash_t +XXH3_len_9to16_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed) +{ + XXH_ASSERT(input != NULL); + XXH_ASSERT(secret != NULL); + XXH_ASSERT(9 <= len && len <= 16); + { xxh_u64 const input_lo = XXH_readLE64(input) ^ (XXH_readLE64(secret) + seed); + xxh_u64 const input_hi = XXH_readLE64(input + len - 8) ^ (XXH_readLE64(secret + 8) - seed); + xxh_u64 const acc = len + (input_lo + input_hi) + XXH3_mul128_fold64(input_lo, input_hi); + return XXH3_avalanche(acc); + } +} + +XXH_FORCE_INLINE XXH64_hash_t +XXH3_len_0to16_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed) +{ + XXH_ASSERT(len <= 16); + { if (len > 8) return XXH3_len_9to16_64b(input, len, secret, seed); + if (len >= 4) return XXH3_len_4to8_64b(input, len, secret, seed); + if (len) return XXH3_len_1to3_64b(input, len, secret, seed); + return 0; + } +} + + +/* === Long Keys === */ + +#define STRIPE_LEN 64 +#define XXH_SECRET_CONSUME_RATE 8 /* nb of secret bytes consumed at each accumulation */ +#define ACC_NB (STRIPE_LEN / sizeof(xxh_u64)) + +typedef enum { XXH3_acc_64bits, XXH3_acc_128bits } XXH3_accWidth_e; + +XXH_FORCE_INLINE void +XXH3_accumulate_512( void* XXH_RESTRICT acc, + const void* XXH_RESTRICT input, + const void* XXH_RESTRICT secret, + XXH3_accWidth_e accWidth) +{ +#if (XXH_VECTOR == XXH_AVX2) + + XXH_ASSERT((((size_t)acc) & 31) == 0); + { XXH_ALIGN(32) __m256i* const xacc = (__m256i *) acc; + const __m256i* const xinput = (const __m256i *) input; /* not really aligned, just for ptr arithmetic, and because _mm256_loadu_si256() requires this type */ + const __m256i* const xsecret = (const __m256i *) secret; /* not really aligned, just for ptr arithmetic, and because _mm256_loadu_si256() requires this type */ + + size_t i; + for (i=0; i < STRIPE_LEN/sizeof(__m256i); i++) { + __m256i const data_vec = _mm256_loadu_si256 (xinput+i); + __m256i const key_vec = _mm256_loadu_si256 (xsecret+i); + __m256i const data_key = _mm256_xor_si256 (data_vec, key_vec); /* uint32 dk[8] = {d0+k0, d1+k1, d2+k2, d3+k3, ...} */ + __m256i const product = _mm256_mul_epu32 (data_key, _mm256_shuffle_epi32 (data_key, 0x31)); /* uint64 mul[4] = {dk0*dk1, dk2*dk3, ...} */ + if (accWidth == XXH3_acc_128bits) { + __m256i const data_swap = _mm256_shuffle_epi32(data_vec, _MM_SHUFFLE(1,0,3,2)); + __m256i const sum = _mm256_add_epi64(xacc[i], data_swap); + xacc[i] = _mm256_add_epi64(product, sum); + } else { /* XXH3_acc_64bits */ + __m256i const sum = _mm256_add_epi64(xacc[i], data_vec); + xacc[i] = _mm256_add_epi64(product, sum); + } + } } + +#elif (XXH_VECTOR == XXH_SSE2) + + XXH_ASSERT((((size_t)acc) & 15) == 0); + { XXH_ALIGN(16) __m128i* const xacc = (__m128i *) acc; + const __m128i* const xinput = (const __m128i *) input; /* not really aligned, just for ptr arithmetic, and because _mm_loadu_si128() requires this type */ + const __m128i* const xsecret = (const __m128i *) secret; /* not really aligned, just for ptr arithmetic, and because _mm_loadu_si128() requires this type */ + + size_t i; + for (i=0; i < STRIPE_LEN/sizeof(__m128i); i++) { + __m128i const data_vec = _mm_loadu_si128 (xinput+i); + __m128i const key_vec = _mm_loadu_si128 (xsecret+i); + __m128i const data_key = _mm_xor_si128 (data_vec, key_vec); /* uint32 dk[8] = {d0+k0, d1+k1, d2+k2, d3+k3, ...} */ + __m128i const product = _mm_mul_epu32 (data_key, _mm_shuffle_epi32 (data_key, 0x31)); /* uint64 mul[4] = {dk0*dk1, dk2*dk3, ...} */ + if (accWidth == XXH3_acc_128bits) { + __m128i const data_swap = _mm_shuffle_epi32(data_vec, _MM_SHUFFLE(1,0,3,2)); + __m128i const sum = _mm_add_epi64(xacc[i], data_swap); + xacc[i] = _mm_add_epi64(product, sum); + } else { /* XXH3_acc_64bits */ + __m128i const sum = _mm_add_epi64(xacc[i], data_vec); + xacc[i] = _mm_add_epi64(product, sum); + } + } } + +#elif (XXH_VECTOR == XXH_NEON) + + XXH_ASSERT((((size_t)acc) & 15) == 0); + { + XXH_ALIGN(16) uint64x2_t* const xacc = (uint64x2_t *) acc; + /* We don't use a uint32x4_t pointer because it causes bus errors on ARMv7. */ + uint8_t const* const xinput = (const uint8_t *) input; + uint8_t const* const xsecret = (const uint8_t *) secret; + + size_t i; + for (i=0; i < STRIPE_LEN / sizeof(uint64x2_t); i++) { +#if !defined(__aarch64__) && !defined(__arm64__) && defined(__GNUC__) /* ARM32-specific hack */ + /* vzip on ARMv7 Clang generates a lot of vmovs (technically vorrs) without this. + * vzip on 32-bit ARM NEON will overwrite the original register, and I think that Clang + * assumes I don't want to destroy it and tries to make a copy. This slows down the code + * a lot. + * aarch64 not only uses an entirely different syntax, but it requires three + * instructions... + * ext v1.16B, v0.16B, #8 // select high bits because aarch64 can't address them directly + * zip1 v3.2s, v0.2s, v1.2s // first zip + * zip2 v2.2s, v0.2s, v1.2s // second zip + * ...to do what ARM does in one: + * vzip.32 d0, d1 // Interleave high and low bits and overwrite. */ + + /* data_vec = xsecret[i]; */ + uint8x16_t const data_vec = vld1q_u8(xinput + (i * 16)); + /* key_vec = xsecret[i]; */ + uint8x16_t const key_vec = vld1q_u8(xsecret + (i * 16)); + /* data_key = data_vec ^ key_vec; */ + uint32x4_t data_key; + + if (accWidth == XXH3_acc_64bits) { + /* Add first to prevent register swaps */ + /* xacc[i] += data_vec; */ + xacc[i] = vaddq_u64 (xacc[i], vreinterpretq_u64_u8(data_vec)); + } else { /* XXH3_acc_128bits */ + /* xacc[i] += swap(data_vec); */ + /* can probably be optimized better */ + uint64x2_t const data64 = vreinterpretq_u64_u8(data_vec); + uint64x2_t const swapped= vextq_u64(data64, data64, 1); + xacc[i] = vaddq_u64 (xacc[i], swapped); + } + + data_key = vreinterpretq_u32_u8(veorq_u8(data_vec, key_vec)); + + /* Here's the magic. We use the quirkiness of vzip to shuffle data_key in place. + * shuffle: data_key[0, 1, 2, 3] = data_key[0, 2, 1, 3] */ + __asm__("vzip.32 %e0, %f0" : "+w" (data_key)); + /* xacc[i] += (uint64x2_t) data_key[0, 1] * (uint64x2_t) data_key[2, 3]; */ + xacc[i] = vmlal_u32(xacc[i], vget_low_u32(data_key), vget_high_u32(data_key)); + +#else + /* On aarch64, vshrn/vmovn seems to be equivalent to, if not faster than, the vzip method. */ + + /* data_vec = xsecret[i]; */ + uint8x16_t const data_vec = vld1q_u8(xinput + (i * 16)); + /* key_vec = xsecret[i]; */ + uint8x16_t const key_vec = vld1q_u8(xsecret + (i * 16)); + /* data_key = data_vec ^ key_vec; */ + uint64x2_t const data_key = vreinterpretq_u64_u8(veorq_u8(data_vec, key_vec)); + /* data_key_lo = (uint32x2_t) (data_key & 0xFFFFFFFF); */ + uint32x2_t const data_key_lo = vmovn_u64 (data_key); + /* data_key_hi = (uint32x2_t) (data_key >> 32); */ + uint32x2_t const data_key_hi = vshrn_n_u64 (data_key, 32); + if (accWidth == XXH3_acc_64bits) { + /* xacc[i] += data_vec; */ + xacc[i] = vaddq_u64 (xacc[i], vreinterpretq_u64_u8(data_vec)); + } else { /* XXH3_acc_128bits */ + /* xacc[i] += swap(data_vec); */ + uint64x2_t const data64 = vreinterpretq_u64_u8(data_vec); + uint64x2_t const swapped= vextq_u64(data64, data64, 1); + xacc[i] = vaddq_u64 (xacc[i], swapped); + } + /* xacc[i] += (uint64x2_t) data_key_lo * (uint64x2_t) data_key_hi; */ + xacc[i] = vmlal_u32 (xacc[i], data_key_lo, data_key_hi); + +#endif + } + } + +#elif (XXH_VECTOR == XXH_VSX) + U64x2* const xacc = (U64x2*) acc; /* presumed aligned */ + U64x2 const* const xinput = (U64x2 const*) input; /* no alignment restriction */ + U64x2 const* const xsecret = (U64x2 const*) secret; /* no alignment restriction */ + U64x2 const v32 = { 32, 32 }; +#if XXH_VSX_BE + U8x16 const vXorSwap = { 0x07, 0x16, 0x25, 0x34, 0x43, 0x52, 0x61, 0x70, + 0x8F, 0x9E, 0xAD, 0xBC, 0xCB, 0xDA, 0xE9, 0xF8 }; +#endif + size_t i; + for (i = 0; i < STRIPE_LEN / sizeof(U64x2); i++) { + /* data_vec = xinput[i]; */ + /* key_vec = xsecret[i]; */ +#if XXH_VSX_BE + /* byteswap */ + U64x2 const data_vec = XXH_vec_revb(vec_vsx_ld(0, xinput + i)); + U64x2 const key_raw = vec_vsx_ld(0, xsecret + i); + /* See comment above. data_key = data_vec ^ swap(xsecret[i]); */ + U64x2 const data_key = (U64x2)XXH_vec_permxor((U8x16)data_vec, (U8x16)key_raw, vXorSwap); +#else + U64x2 const data_vec = vec_vsx_ld(0, xinput + i); + U64x2 const key_vec = vec_vsx_ld(0, xsecret + i); + U64x2 const data_key = data_vec ^ key_vec; +#endif + /* shuffled = (data_key << 32) | (data_key >> 32); */ + U32x4 const shuffled = (U32x4)vec_rl(data_key, v32); + /* product = ((U64x2)data_key & 0xFFFFFFFF) * ((U64x2)shuffled & 0xFFFFFFFF); */ + U64x2 const product = XXH_vec_mulo((U32x4)data_key, shuffled); + xacc[i] += product; + + if (accWidth == XXH3_acc_64bits) { + xacc[i] += data_vec; + } else { /* XXH3_acc_128bits */ + /* swap high and low halves */ + U64x2 const data_swapped = vec_xxpermdi(data_vec, data_vec, 2); + xacc[i] += data_swapped; + } + } + +#else /* scalar variant of Accumulator - universal */ + + XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64* const xacc = (xxh_u64*) acc; /* presumed aligned on 32-bytes boundaries, little hint for the auto-vectorizer */ + const xxh_u8* const xinput = (const xxh_u8*) input; /* no alignment restriction */ + const xxh_u8* const xsecret = (const xxh_u8*) secret; /* no alignment restriction */ + size_t i; + XXH_ASSERT(((size_t)acc & (XXH_ACC_ALIGN-1)) == 0); + for (i=0; i < ACC_NB; i++) { + xxh_u64 const data_val = XXH_readLE64(xinput + 8*i); + xxh_u64 const data_key = data_val ^ XXH_readLE64(xsecret + i*8); + + if (accWidth == XXH3_acc_64bits) { + xacc[i] += data_val; + } else { + xacc[i ^ 1] += data_val; /* swap adjacent lanes */ + } + xacc[i] += XXH_mult32to64(data_key & 0xFFFFFFFF, data_key >> 32); + } +#endif +} + +XXH_FORCE_INLINE void +XXH3_scrambleAcc(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret) +{ +#if (XXH_VECTOR == XXH_AVX2) + + XXH_ASSERT((((size_t)acc) & 31) == 0); + { XXH_ALIGN(32) __m256i* const xacc = (__m256i*) acc; + const __m256i* const xsecret = (const __m256i *) secret; /* not really aligned, just for ptr arithmetic, and because _mm256_loadu_si256() requires this argument type */ + const __m256i prime32 = _mm256_set1_epi32((int)PRIME32_1); + + size_t i; + for (i=0; i < STRIPE_LEN/sizeof(__m256i); i++) { + /* xacc[i] ^= (xacc[i] >> 47) */ + __m256i const acc_vec = xacc[i]; + __m256i const shifted = _mm256_srli_epi64 (acc_vec, 47); + __m256i const data_vec = _mm256_xor_si256 (acc_vec, shifted); + /* xacc[i] ^= xsecret; */ + __m256i const key_vec = _mm256_loadu_si256 (xsecret+i); + __m256i const data_key = _mm256_xor_si256 (data_vec, key_vec); + + /* xacc[i] *= PRIME32_1; */ + __m256i const data_key_hi = _mm256_shuffle_epi32 (data_key, 0x31); + __m256i const prod_lo = _mm256_mul_epu32 (data_key, prime32); + __m256i const prod_hi = _mm256_mul_epu32 (data_key_hi, prime32); + xacc[i] = _mm256_add_epi64(prod_lo, _mm256_slli_epi64(prod_hi, 32)); + } + } + +#elif (XXH_VECTOR == XXH_SSE2) + + XXH_ASSERT((((size_t)acc) & 15) == 0); + { XXH_ALIGN(16) __m128i* const xacc = (__m128i*) acc; + const __m128i* const xsecret = (const __m128i *) secret; /* not really aligned, just for ptr arithmetic, and because _mm_loadu_si128() requires this argument type */ + const __m128i prime32 = _mm_set1_epi32((int)PRIME32_1); + + size_t i; + for (i=0; i < STRIPE_LEN/sizeof(__m128i); i++) { + /* xacc[i] ^= (xacc[i] >> 47) */ + __m128i const acc_vec = xacc[i]; + __m128i const shifted = _mm_srli_epi64 (acc_vec, 47); + __m128i const data_vec = _mm_xor_si128 (acc_vec, shifted); + /* xacc[i] ^= xsecret; */ + __m128i const key_vec = _mm_loadu_si128 (xsecret+i); + __m128i const data_key = _mm_xor_si128 (data_vec, key_vec); + + /* xacc[i] *= PRIME32_1; */ + __m128i const data_key_hi = _mm_shuffle_epi32 (data_key, 0x31); + __m128i const prod_lo = _mm_mul_epu32 (data_key, prime32); + __m128i const prod_hi = _mm_mul_epu32 (data_key_hi, prime32); + xacc[i] = _mm_add_epi64(prod_lo, _mm_slli_epi64(prod_hi, 32)); + } + } + +#elif (XXH_VECTOR == XXH_NEON) + + XXH_ASSERT((((size_t)acc) & 15) == 0); + + { uint64x2_t* const xacc = (uint64x2_t*) acc; + uint8_t const* const xsecret = (uint8_t const*) secret; + uint32x2_t const prime = vdup_n_u32 (PRIME32_1); + + size_t i; + for (i=0; i < STRIPE_LEN/sizeof(uint64x2_t); i++) { + /* data_vec = xacc[i] ^ (xacc[i] >> 47); */ + uint64x2_t const acc_vec = xacc[i]; + uint64x2_t const shifted = vshrq_n_u64 (acc_vec, 47); + uint64x2_t const data_vec = veorq_u64 (acc_vec, shifted); + + /* key_vec = xsecret[i]; */ + uint32x4_t const key_vec = vreinterpretq_u32_u8(vld1q_u8(xsecret + (i * 16))); + /* data_key = data_vec ^ key_vec; */ + uint32x4_t const data_key = veorq_u32 (vreinterpretq_u32_u64(data_vec), key_vec); + /* shuffled = { data_key[0, 2], data_key[1, 3] }; */ + uint32x2x2_t const shuffled = vzip_u32 (vget_low_u32(data_key), vget_high_u32(data_key)); + + /* data_key *= PRIME32_1 */ + + /* prod_hi = (data_key >> 32) * PRIME32_1; */ + uint64x2_t const prod_hi = vmull_u32 (shuffled.val[1], prime); + /* xacc[i] = prod_hi << 32; */ + xacc[i] = vshlq_n_u64(prod_hi, 32); + /* xacc[i] += (prod_hi & 0xFFFFFFFF) * PRIME32_1; */ + xacc[i] = vmlal_u32(xacc[i], shuffled.val[0], prime); + } } + +#elif (XXH_VECTOR == XXH_VSX) + + U64x2* const xacc = (U64x2*) acc; + const U64x2* const xsecret = (const U64x2*) secret; + /* constants */ + U64x2 const v32 = { 32, 32 }; + U64x2 const v47 = { 47, 47 }; + U32x4 const prime = { PRIME32_1, PRIME32_1, PRIME32_1, PRIME32_1 }; + size_t i; +#if XXH_VSX_BE + /* endian swap */ + U8x16 const vXorSwap = { 0x07, 0x16, 0x25, 0x34, 0x43, 0x52, 0x61, 0x70, + 0x8F, 0x9E, 0xAD, 0xBC, 0xCB, 0xDA, 0xE9, 0xF8 }; +#endif + for (i = 0; i < STRIPE_LEN / sizeof(U64x2); i++) { + U64x2 const acc_vec = xacc[i]; + U64x2 const data_vec = acc_vec ^ (acc_vec >> v47); + /* key_vec = xsecret[i]; */ +#if XXH_VSX_BE + /* swap bytes words */ + U64x2 const key_raw = vec_vsx_ld(0, xsecret + i); + U64x2 const data_key = (U64x2)XXH_vec_permxor((U8x16)data_vec, (U8x16)key_raw, vXorSwap); +#else + U64x2 const key_vec = vec_vsx_ld(0, xsecret + i); + U64x2 const data_key = data_vec ^ key_vec; +#endif + + /* data_key *= PRIME32_1 */ + + /* prod_lo = ((U64x2)data_key & 0xFFFFFFFF) * ((U64x2)prime & 0xFFFFFFFF); */ + U64x2 const prod_even = XXH_vec_mule((U32x4)data_key, prime); + /* prod_hi = ((U64x2)data_key >> 32) * ((U64x2)prime >> 32); */ + U64x2 const prod_odd = XXH_vec_mulo((U32x4)data_key, prime); + xacc[i] = prod_odd + (prod_even << v32); + } + +#else /* scalar variant of Scrambler - universal */ + + XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64* const xacc = (xxh_u64*) acc; /* presumed aligned on 32-bytes boundaries, little hint for the auto-vectorizer */ + const xxh_u8* const xsecret = (const xxh_u8*) secret; /* no alignment restriction */ + size_t i; + XXH_ASSERT((((size_t)acc) & (XXH_ACC_ALIGN-1)) == 0); + for (i=0; i < ACC_NB; i++) { + xxh_u64 const key64 = XXH_readLE64(xsecret + 8*i); + xxh_u64 acc64 = xacc[i]; + acc64 ^= acc64 >> 47; + acc64 ^= key64; + acc64 *= PRIME32_1; + xacc[i] = acc64; + } + +#endif +} + +/* assumption : nbStripes will not overflow secret size */ +XXH_FORCE_INLINE void +XXH3_accumulate( xxh_u64* XXH_RESTRICT acc, + const xxh_u8* XXH_RESTRICT input, + const xxh_u8* XXH_RESTRICT secret, + size_t nbStripes, + XXH3_accWidth_e accWidth) +{ + size_t n; + for (n = 0; n < nbStripes; n++ ) { + XXH3_accumulate_512(acc, + input + n*STRIPE_LEN, + secret + n*XXH_SECRET_CONSUME_RATE, + accWidth); + } +} + +/* note : clang auto-vectorizes well in SS2 mode _if_ this function is `static`, + * and doesn't auto-vectorize it at all if it is `FORCE_INLINE`. + * However, it auto-vectorizes better AVX2 if it is `FORCE_INLINE` + * Pretty much every other modes and compilers prefer `FORCE_INLINE`. + */ + +#if defined(__clang__) && (XXH_VECTOR==0) && !defined(__AVX2__) && !defined(__arm__) && !defined(__thumb__) +static void +#else +XXH_FORCE_INLINE void +#endif +XXH3_hashLong_internal_loop( xxh_u64* XXH_RESTRICT acc, + const xxh_u8* XXH_RESTRICT input, size_t len, + const xxh_u8* XXH_RESTRICT secret, size_t secretSize, + XXH3_accWidth_e accWidth) +{ + size_t const nb_rounds = (secretSize - STRIPE_LEN) / XXH_SECRET_CONSUME_RATE; + size_t const block_len = STRIPE_LEN * nb_rounds; + size_t const nb_blocks = len / block_len; + + size_t n; + + XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); + + for (n = 0; n < nb_blocks; n++) { + XXH3_accumulate(acc, input + n*block_len, secret, nb_rounds, accWidth); + XXH3_scrambleAcc(acc, secret + secretSize - STRIPE_LEN); + } + + /* last partial block */ + XXH_ASSERT(len > STRIPE_LEN); + { size_t const nbStripes = (len - (block_len * nb_blocks)) / STRIPE_LEN; + XXH_ASSERT(nbStripes <= (secretSize / XXH_SECRET_CONSUME_RATE)); + XXH3_accumulate(acc, input + nb_blocks*block_len, secret, nbStripes, accWidth); + + /* last stripe */ + if (len & (STRIPE_LEN - 1)) { + const xxh_u8* const p = input + len - STRIPE_LEN; +#define XXH_SECRET_LASTACC_START 7 /* do not align on 8, so that secret is different from scrambler */ + XXH3_accumulate_512(acc, p, secret + secretSize - STRIPE_LEN - XXH_SECRET_LASTACC_START, accWidth); + } } +} + +XXH_FORCE_INLINE xxh_u64 +XXH3_mix2Accs(const xxh_u64* XXH_RESTRICT acc, const xxh_u8* XXH_RESTRICT secret) +{ + return XXH3_mul128_fold64( + acc[0] ^ XXH_readLE64(secret), + acc[1] ^ XXH_readLE64(secret+8) ); +} + +static XXH64_hash_t +XXH3_mergeAccs(const xxh_u64* XXH_RESTRICT acc, const xxh_u8* XXH_RESTRICT secret, xxh_u64 start) +{ + xxh_u64 result64 = start; + + result64 += XXH3_mix2Accs(acc+0, secret + 0); + result64 += XXH3_mix2Accs(acc+2, secret + 16); + result64 += XXH3_mix2Accs(acc+4, secret + 32); + result64 += XXH3_mix2Accs(acc+6, secret + 48); + + return XXH3_avalanche(result64); +} + +#define XXH3_INIT_ACC { PRIME32_3, PRIME64_1, PRIME64_2, PRIME64_3, \ + PRIME64_4, PRIME32_2, PRIME64_5, PRIME32_1 }; + +XXH_FORCE_INLINE XXH64_hash_t +XXH3_hashLong_internal(const xxh_u8* XXH_RESTRICT input, size_t len, + const xxh_u8* XXH_RESTRICT secret, size_t secretSize) +{ + XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[ACC_NB] = XXH3_INIT_ACC; + + XXH3_hashLong_internal_loop(acc, input, len, secret, secretSize, XXH3_acc_64bits); + + /* converge into final hash */ + XXH_STATIC_ASSERT(sizeof(acc) == 64); +#define XXH_SECRET_MERGEACCS_START 11 /* do not align on 8, so that secret is different from accumulator */ + XXH_ASSERT(secretSize >= sizeof(acc) + XXH_SECRET_MERGEACCS_START); + return XXH3_mergeAccs(acc, secret + XXH_SECRET_MERGEACCS_START, (xxh_u64)len * PRIME64_1); +} + + +XXH_NO_INLINE XXH64_hash_t /* It's important for performance that XXH3_hashLong is not inlined. Not sure why (uop cache maybe ?), but difference is large and easily measurable */ +XXH3_hashLong_64b_defaultSecret(const xxh_u8* XXH_RESTRICT input, size_t len) +{ + return XXH3_hashLong_internal(input, len, kSecret, sizeof(kSecret)); +} + +XXH_NO_INLINE XXH64_hash_t /* It's important for performance that XXH3_hashLong is not inlined. Not sure why (uop cache maybe ?), but difference is large and easily measurable */ +XXH3_hashLong_64b_withSecret(const xxh_u8* XXH_RESTRICT input, size_t len, + const xxh_u8* XXH_RESTRICT secret, size_t secretSize) +{ + return XXH3_hashLong_internal(input, len, secret, secretSize); +} + + +XXH_FORCE_INLINE void XXH_writeLE64(void* dst, xxh_u64 v64) +{ + if (!XXH_CPU_LITTLE_ENDIAN) v64 = XXH_swap64(v64); + memcpy(dst, &v64, sizeof(v64)); +} + +/* XXH3_initCustomSecret() : + * destination `customSecret` is presumed allocated and same size as `kSecret`. + */ +XXH_FORCE_INLINE void XXH3_initCustomSecret(xxh_u8* customSecret, xxh_u64 seed64) +{ + int const nbRounds = XXH_SECRET_DEFAULT_SIZE / 16; + int i; + + XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 15) == 0); + + for (i=0; i < nbRounds; i++) { + XXH_writeLE64(customSecret + 16*i, XXH_readLE64(kSecret + 16*i) + seed64); + XXH_writeLE64(customSecret + 16*i + 8, XXH_readLE64(kSecret + 16*i + 8) - seed64); + } +} + + +/* XXH3_hashLong_64b_withSeed() : + * Generate a custom key, + * based on alteration of default kSecret with the seed, + * and then use this key for long mode hashing. + * This operation is decently fast but nonetheless costs a little bit of time. + * Try to avoid it whenever possible (typically when seed==0). + */ +XXH_NO_INLINE XXH64_hash_t /* It's important for performance that XXH3_hashLong is not inlined. Not sure why (uop cache maybe ?), but difference is large and easily measurable */ +XXH3_hashLong_64b_withSeed(const xxh_u8* input, size_t len, XXH64_hash_t seed) +{ + XXH_ALIGN(8) xxh_u8 secret[XXH_SECRET_DEFAULT_SIZE]; + if (seed==0) return XXH3_hashLong_64b_defaultSecret(input, len); + XXH3_initCustomSecret(secret, seed); + return XXH3_hashLong_internal(input, len, secret, sizeof(secret)); +} + + +XXH_FORCE_INLINE xxh_u64 XXH3_mix16B(const xxh_u8* XXH_RESTRICT input, + const xxh_u8* XXH_RESTRICT secret, xxh_u64 seed64) +{ + xxh_u64 const input_lo = XXH_readLE64(input); + xxh_u64 const input_hi = XXH_readLE64(input+8); + return XXH3_mul128_fold64( + input_lo ^ (XXH_readLE64(secret) + seed64), + input_hi ^ (XXH_readLE64(secret+8) - seed64) ); +} + + +XXH_FORCE_INLINE XXH64_hash_t +XXH3_len_17to128_64b(const xxh_u8* XXH_RESTRICT input, size_t len, + const xxh_u8* XXH_RESTRICT secret, size_t secretSize, + XXH64_hash_t seed) +{ + XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize; + XXH_ASSERT(16 < len && len <= 128); + + { xxh_u64 acc = len * PRIME64_1; + if (len > 32) { + if (len > 64) { + if (len > 96) { + acc += XXH3_mix16B(input+48, secret+96, seed); + acc += XXH3_mix16B(input+len-64, secret+112, seed); + } + acc += XXH3_mix16B(input+32, secret+64, seed); + acc += XXH3_mix16B(input+len-48, secret+80, seed); + } + acc += XXH3_mix16B(input+16, secret+32, seed); + acc += XXH3_mix16B(input+len-32, secret+48, seed); + } + acc += XXH3_mix16B(input+0, secret+0, seed); + acc += XXH3_mix16B(input+len-16, secret+16, seed); + + return XXH3_avalanche(acc); + } +} + +#define XXH3_MIDSIZE_MAX 240 + +XXH_NO_INLINE XXH64_hash_t +XXH3_len_129to240_64b(const xxh_u8* XXH_RESTRICT input, size_t len, + const xxh_u8* XXH_RESTRICT secret, size_t secretSize, + XXH64_hash_t seed) +{ + XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize; + XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX); + + #define XXH3_MIDSIZE_STARTOFFSET 3 + #define XXH3_MIDSIZE_LASTOFFSET 17 + + { xxh_u64 acc = len * PRIME64_1; + int const nbRounds = (int)len / 16; + int i; + for (i=0; i<8; i++) { + acc += XXH3_mix16B(input+(16*i), secret+(16*i), seed); + } + acc = XXH3_avalanche(acc); + XXH_ASSERT(nbRounds >= 8); + for (i=8 ; i < nbRounds; i++) { + acc += XXH3_mix16B(input+(16*i), secret+(16*(i-8)) + XXH3_MIDSIZE_STARTOFFSET, seed); + } + /* last bytes */ + acc += XXH3_mix16B(input + len - 16, secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET, seed); + return XXH3_avalanche(acc); + } +} + +/* === Public entry point === */ + +XXH_PUBLIC_API XXH64_hash_t XXH3_64bits(const void* input, size_t len) +{ + if (len <= 16) return XXH3_len_0to16_64b((const xxh_u8*)input, len, kSecret, 0); + if (len <= 128) return XXH3_len_17to128_64b((const xxh_u8*)input, len, kSecret, sizeof(kSecret), 0); + if (len <= XXH3_MIDSIZE_MAX) return XXH3_len_129to240_64b((const xxh_u8*)input, len, kSecret, sizeof(kSecret), 0); + return XXH3_hashLong_64b_defaultSecret((const xxh_u8*)input, len); +} + +XXH_PUBLIC_API XXH64_hash_t +XXH3_64bits_withSecret(const void* input, size_t len, const void* secret, size_t secretSize) +{ + XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); + /* if an action must be taken should `secret` conditions not be respected, + * it should be done here. + * For now, it's a contract pre-condition. + * Adding a check and a branch here would cost performance at every hash */ + if (len <= 16) return XXH3_len_0to16_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, 0); + if (len <= 128) return XXH3_len_17to128_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretSize, 0); + if (len <= XXH3_MIDSIZE_MAX) return XXH3_len_129to240_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretSize, 0); + return XXH3_hashLong_64b_withSecret((const xxh_u8*)input, len, (const xxh_u8*)secret, secretSize); +} + +XXH_PUBLIC_API XXH64_hash_t +XXH3_64bits_withSeed(const void* input, size_t len, XXH64_hash_t seed) +{ + if (len <= 16) return XXH3_len_0to16_64b((const xxh_u8*)input, len, kSecret, seed); + if (len <= 128) return XXH3_len_17to128_64b((const xxh_u8*)input, len, kSecret, sizeof(kSecret), seed); + if (len <= XXH3_MIDSIZE_MAX) return XXH3_len_129to240_64b((const xxh_u8*)input, len, kSecret, sizeof(kSecret), seed); + return XXH3_hashLong_64b_withSeed((const xxh_u8*)input, len, seed); +} + +/* === XXH3 streaming === */ + +XXH_PUBLIC_API XXH3_state_t* XXH3_createState(void) +{ + return (XXH3_state_t*)XXH_malloc(sizeof(XXH3_state_t)); +} + +XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t* statePtr) +{ + XXH_free(statePtr); + return XXH_OK; +} + +XXH_PUBLIC_API void +XXH3_copyState(XXH3_state_t* dst_state, const XXH3_state_t* src_state) +{ + memcpy(dst_state, src_state, sizeof(*dst_state)); +} + +static void +XXH3_64bits_reset_internal(XXH3_state_t* statePtr, + XXH64_hash_t seed, + const xxh_u8* secret, size_t secretSize) +{ + XXH_ASSERT(statePtr != NULL); + memset(statePtr, 0, sizeof(*statePtr)); + statePtr->acc[0] = PRIME32_3; + statePtr->acc[1] = PRIME64_1; + statePtr->acc[2] = PRIME64_2; + statePtr->acc[3] = PRIME64_3; + statePtr->acc[4] = PRIME64_4; + statePtr->acc[5] = PRIME32_2; + statePtr->acc[6] = PRIME64_5; + statePtr->acc[7] = PRIME32_1; + statePtr->seed = seed; + XXH_ASSERT(secret != NULL); + statePtr->secret = secret; + XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); + statePtr->secretLimit = (XXH32_hash_t)(secretSize - STRIPE_LEN); + statePtr->nbStripesPerBlock = statePtr->secretLimit / XXH_SECRET_CONSUME_RATE; +} + +XXH_PUBLIC_API XXH_errorcode +XXH3_64bits_reset(XXH3_state_t* statePtr) +{ + if (statePtr == NULL) return XXH_ERROR; + XXH3_64bits_reset_internal(statePtr, 0, kSecret, XXH_SECRET_DEFAULT_SIZE); + return XXH_OK; +} + +XXH_PUBLIC_API XXH_errorcode +XXH3_64bits_reset_withSecret(XXH3_state_t* statePtr, const void* secret, size_t secretSize) +{ + if (statePtr == NULL) return XXH_ERROR; + XXH3_64bits_reset_internal(statePtr, 0, (const xxh_u8*)secret, secretSize); + if (secret == NULL) return XXH_ERROR; + if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR; + return XXH_OK; +} + +XXH_PUBLIC_API XXH_errorcode +XXH3_64bits_reset_withSeed(XXH3_state_t* statePtr, XXH64_hash_t seed) +{ + if (statePtr == NULL) return XXH_ERROR; + XXH3_64bits_reset_internal(statePtr, seed, kSecret, XXH_SECRET_DEFAULT_SIZE); + XXH3_initCustomSecret(statePtr->customSecret, seed); + statePtr->secret = statePtr->customSecret; + return XXH_OK; +} + +XXH_FORCE_INLINE void +XXH3_consumeStripes( xxh_u64* acc, + XXH32_hash_t* nbStripesSoFarPtr, XXH32_hash_t nbStripesPerBlock, + const xxh_u8* input, size_t totalStripes, + const xxh_u8* secret, size_t secretLimit, + XXH3_accWidth_e accWidth) +{ + XXH_ASSERT(*nbStripesSoFarPtr < nbStripesPerBlock); + if (nbStripesPerBlock - *nbStripesSoFarPtr <= totalStripes) { + /* need a scrambling operation */ + size_t const nbStripes = nbStripesPerBlock - *nbStripesSoFarPtr; + XXH3_accumulate(acc, input, secret + nbStripesSoFarPtr[0] * XXH_SECRET_CONSUME_RATE, nbStripes, accWidth); + XXH3_scrambleAcc(acc, secret + secretLimit); + XXH3_accumulate(acc, input + nbStripes * STRIPE_LEN, secret, totalStripes - nbStripes, accWidth); + *nbStripesSoFarPtr = (XXH32_hash_t)(totalStripes - nbStripes); + } else { + XXH3_accumulate(acc, input, secret + nbStripesSoFarPtr[0] * XXH_SECRET_CONSUME_RATE, totalStripes, accWidth); + *nbStripesSoFarPtr += (XXH32_hash_t)totalStripes; + } +} + +XXH_FORCE_INLINE XXH_errorcode +XXH3_update(XXH3_state_t* state, const xxh_u8* input, size_t len, XXH3_accWidth_e accWidth) +{ + if (input==NULL) +#if defined(XXH_ACCEPT_NULL_INPUT_POINTER) && (XXH_ACCEPT_NULL_INPUT_POINTER>=1) + return XXH_OK; +#else + return XXH_ERROR; +#endif + + { const xxh_u8* const bEnd = input + len; + + state->totalLen += len; + + if (state->bufferedSize + len <= XXH3_INTERNALBUFFER_SIZE) { /* fill in tmp buffer */ + XXH_memcpy(state->buffer + state->bufferedSize, input, len); + state->bufferedSize += (XXH32_hash_t)len; + return XXH_OK; + } + /* input now > XXH3_INTERNALBUFFER_SIZE */ + + #define XXH3_INTERNALBUFFER_STRIPES (XXH3_INTERNALBUFFER_SIZE / STRIPE_LEN) + XXH_STATIC_ASSERT(XXH3_INTERNALBUFFER_SIZE % STRIPE_LEN == 0); /* clean multiple */ + + if (state->bufferedSize) { /* some input within internal buffer: fill then consume it */ + size_t const loadSize = XXH3_INTERNALBUFFER_SIZE - state->bufferedSize; + XXH_memcpy(state->buffer + state->bufferedSize, input, loadSize); + input += loadSize; + XXH3_consumeStripes(state->acc, + &state->nbStripesSoFar, state->nbStripesPerBlock, + state->buffer, XXH3_INTERNALBUFFER_STRIPES, + state->secret, state->secretLimit, + accWidth); + state->bufferedSize = 0; + } + + /* consume input by full buffer quantities */ + if (input+XXH3_INTERNALBUFFER_SIZE <= bEnd) { + const xxh_u8* const limit = bEnd - XXH3_INTERNALBUFFER_SIZE; + do { + XXH3_consumeStripes(state->acc, + &state->nbStripesSoFar, state->nbStripesPerBlock, + input, XXH3_INTERNALBUFFER_STRIPES, + state->secret, state->secretLimit, + accWidth); + input += XXH3_INTERNALBUFFER_SIZE; + } while (input<=limit); + } + + if (input < bEnd) { /* some remaining input input : buffer it */ + XXH_memcpy(state->buffer, input, (size_t)(bEnd-input)); + state->bufferedSize = (XXH32_hash_t)(bEnd-input); + } + } + + return XXH_OK; +} + +XXH_PUBLIC_API XXH_errorcode +XXH3_64bits_update(XXH3_state_t* state, const void* input, size_t len) +{ + return XXH3_update(state, (const xxh_u8*)input, len, XXH3_acc_64bits); +} + + +XXH_FORCE_INLINE void +XXH3_digest_long (XXH64_hash_t* acc, const XXH3_state_t* state, XXH3_accWidth_e accWidth) +{ + memcpy(acc, state->acc, sizeof(state->acc)); /* digest locally, state remains unaltered, and can continue ingesting more input afterwards */ + if (state->bufferedSize >= STRIPE_LEN) { + size_t const totalNbStripes = state->bufferedSize / STRIPE_LEN; + XXH32_hash_t nbStripesSoFar = state->nbStripesSoFar; + XXH3_consumeStripes(acc, + &nbStripesSoFar, state->nbStripesPerBlock, + state->buffer, totalNbStripes, + state->secret, state->secretLimit, + accWidth); + if (state->bufferedSize % STRIPE_LEN) { /* one last partial stripe */ + XXH3_accumulate_512(acc, + state->buffer + state->bufferedSize - STRIPE_LEN, + state->secret + state->secretLimit - XXH_SECRET_LASTACC_START, + accWidth); + } + } else { /* bufferedSize < STRIPE_LEN */ + if (state->bufferedSize) { /* one last stripe */ + xxh_u8 lastStripe[STRIPE_LEN]; + size_t const catchupSize = STRIPE_LEN - state->bufferedSize; + memcpy(lastStripe, state->buffer + sizeof(state->buffer) - catchupSize, catchupSize); + memcpy(lastStripe + catchupSize, state->buffer, state->bufferedSize); + XXH3_accumulate_512(acc, + lastStripe, + state->secret + state->secretLimit - XXH_SECRET_LASTACC_START, + accWidth); + } } +} + +XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_digest (const XXH3_state_t* state) +{ + if (state->totalLen > XXH3_MIDSIZE_MAX) { + XXH_ALIGN(XXH_ACC_ALIGN) XXH64_hash_t acc[ACC_NB]; + XXH3_digest_long(acc, state, XXH3_acc_64bits); + return XXH3_mergeAccs(acc, state->secret + XXH_SECRET_MERGEACCS_START, (xxh_u64)state->totalLen * PRIME64_1); + } + /* len <= XXH3_MIDSIZE_MAX : short code */ + if (state->seed) + return XXH3_64bits_withSeed(state->buffer, (size_t)state->totalLen, state->seed); + return XXH3_64bits_withSecret(state->buffer, (size_t)(state->totalLen), state->secret, state->secretLimit + STRIPE_LEN); +} + +/* ========================================== + * XXH3 128 bits (=> XXH128) + * ========================================== */ + +XXH_FORCE_INLINE XXH128_hash_t +XXH3_len_1to3_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed) +{ + XXH_ASSERT(input != NULL); + XXH_ASSERT(1 <= len && len <= 3); + XXH_ASSERT(secret != NULL); + { xxh_u8 const c1 = input[0]; + xxh_u8 const c2 = input[len >> 1]; + xxh_u8 const c3 = input[len - 1]; + xxh_u32 const combinedl = ((xxh_u32)c1) + (((xxh_u32)c2) << 8) + (((xxh_u32)c3) << 16) + (((xxh_u32)len) << 24); + xxh_u32 const combinedh = XXH_swap32(combinedl); + xxh_u64 const keyed_lo = (xxh_u64)combinedl ^ (XXH_readLE32(secret) + seed); + xxh_u64 const keyed_hi = (xxh_u64)combinedh ^ (XXH_readLE32(secret+4) - seed); + xxh_u64 const mixedl = keyed_lo * PRIME64_1; + xxh_u64 const mixedh = keyed_hi * PRIME64_5; + XXH128_hash_t const h128 = { XXH3_avalanche(mixedl) /*low64*/, XXH3_avalanche(mixedh) /*high64*/ }; + return h128; + } +} + + +XXH_FORCE_INLINE XXH128_hash_t +XXH3_len_4to8_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed) +{ + XXH_ASSERT(input != NULL); + XXH_ASSERT(secret != NULL); + XXH_ASSERT(4 <= len && len <= 8); + { xxh_u32 const input_lo = XXH_readLE32(input); + xxh_u32 const input_hi = XXH_readLE32(input + len - 4); + xxh_u64 const input_64_lo = input_lo + ((xxh_u64)input_hi << 32); + xxh_u64 const input_64_hi = XXH_swap64(input_64_lo); + xxh_u64 const keyed_lo = input_64_lo ^ (XXH_readLE64(secret) + seed); + xxh_u64 const keyed_hi = input_64_hi ^ (XXH_readLE64(secret + 8) - seed); + xxh_u64 const mix64l1 = len + ((keyed_lo ^ (keyed_lo >> 51)) * PRIME32_1); + xxh_u64 const mix64l2 = (mix64l1 ^ (mix64l1 >> 47)) * PRIME64_2; + xxh_u64 const mix64h1 = ((keyed_hi ^ (keyed_hi >> 47)) * PRIME64_1) - len; + xxh_u64 const mix64h2 = (mix64h1 ^ (mix64h1 >> 43)) * PRIME64_4; + { XXH128_hash_t const h128 = { XXH3_avalanche(mix64l2) /*low64*/, XXH3_avalanche(mix64h2) /*high64*/ }; + return h128; + } } +} + +XXH_FORCE_INLINE XXH128_hash_t +XXH3_len_9to16_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed) +{ + XXH_ASSERT(input != NULL); + XXH_ASSERT(secret != NULL); + XXH_ASSERT(9 <= len && len <= 16); + { xxh_u64 const input_lo = XXH_readLE64(input) ^ (XXH_readLE64(secret) + seed); + xxh_u64 const input_hi = XXH_readLE64(input + len - 8) ^ (XXH_readLE64(secret+8) - seed); + XXH128_hash_t m128 = XXH_mult64to128(input_lo ^ input_hi, PRIME64_1); + xxh_u64 const lenContrib = XXH_mult32to64(len, PRIME32_5); + m128.low64 += lenContrib; + m128.high64 += input_hi * PRIME64_1; + m128.low64 ^= (m128.high64 >> 32); + { XXH128_hash_t h128 = XXH_mult64to128(m128.low64, PRIME64_2); + h128.high64 += m128.high64 * PRIME64_2; + h128.low64 = XXH3_avalanche(h128.low64); + h128.high64 = XXH3_avalanche(h128.high64); + return h128; + } } +} + +/* Assumption : `secret` size is >= 16 + * Note : it should be >= XXH3_SECRET_SIZE_MIN anyway */ +XXH_FORCE_INLINE XXH128_hash_t +XXH3_len_0to16_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed) +{ + XXH_ASSERT(len <= 16); + { if (len > 8) return XXH3_len_9to16_128b(input, len, secret, seed); + if (len >= 4) return XXH3_len_4to8_128b(input, len, secret, seed); + if (len) return XXH3_len_1to3_128b(input, len, secret, seed); + { XXH128_hash_t const h128 = { 0, 0 }; + return h128; + } } +} + +XXH_FORCE_INLINE XXH128_hash_t +XXH3_hashLong_128b_internal(const xxh_u8* XXH_RESTRICT input, size_t len, + const xxh_u8* XXH_RESTRICT secret, size_t secretSize) +{ + XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[ACC_NB] = XXH3_INIT_ACC; + + XXH3_hashLong_internal_loop(acc, input, len, secret, secretSize, XXH3_acc_128bits); + + /* converge into final hash */ + XXH_STATIC_ASSERT(sizeof(acc) == 64); + XXH_ASSERT(secretSize >= sizeof(acc) + XXH_SECRET_MERGEACCS_START); + { xxh_u64 const low64 = XXH3_mergeAccs(acc, secret + XXH_SECRET_MERGEACCS_START, (xxh_u64)len * PRIME64_1); + xxh_u64 const high64 = XXH3_mergeAccs(acc, secret + secretSize - sizeof(acc) - XXH_SECRET_MERGEACCS_START, ~((xxh_u64)len * PRIME64_2)); + XXH128_hash_t const h128 = { low64, high64 }; + return h128; + } +} + +XXH_NO_INLINE XXH128_hash_t /* It's important for performance that XXH3_hashLong is not inlined. Not sure why (uop cache maybe ?), but difference is large and easily measurable */ +XXH3_hashLong_128b_defaultSecret(const xxh_u8* input, size_t len) +{ + return XXH3_hashLong_128b_internal(input, len, kSecret, sizeof(kSecret)); +} + +XXH_NO_INLINE XXH128_hash_t /* It's important for performance that XXH3_hashLong is not inlined. Not sure why (uop cache maybe ?), but difference is large and easily measurable */ +XXH3_hashLong_128b_withSecret(const xxh_u8* input, size_t len, + const xxh_u8* secret, size_t secretSize) +{ + return XXH3_hashLong_128b_internal(input, len, secret, secretSize); +} + +XXH_NO_INLINE XXH128_hash_t /* It's important for performance that XXH3_hashLong is not inlined. Not sure why (uop cache maybe ?), but difference is large and easily measurable */ +XXH3_hashLong_128b_withSeed(const xxh_u8* input, size_t len, XXH64_hash_t seed) +{ + XXH_ALIGN(8) xxh_u8 secret[XXH_SECRET_DEFAULT_SIZE]; + if (seed == 0) return XXH3_hashLong_128b_defaultSecret(input, len); + XXH3_initCustomSecret(secret, seed); + return XXH3_hashLong_128b_internal(input, len, secret, sizeof(secret)); +} + + +XXH_FORCE_INLINE XXH128_hash_t +XXH128_mix32B(XXH128_hash_t acc, const xxh_u8* input_1, const xxh_u8* input_2, const xxh_u8* secret, XXH64_hash_t seed) +{ + acc.low64 += XXH3_mix16B (input_1, secret+0, seed); + acc.low64 ^= XXH_readLE64(input_2) + XXH_readLE64(input_2 + 8); + acc.high64 += XXH3_mix16B (input_2, secret+16, seed); + acc.high64 ^= XXH_readLE64(input_1) + XXH_readLE64(input_1 + 8); + return acc; +} + +XXH_NO_INLINE XXH128_hash_t +XXH3_len_129to240_128b(const xxh_u8* XXH_RESTRICT input, size_t len, + const xxh_u8* XXH_RESTRICT secret, size_t secretSize, + XXH64_hash_t seed) +{ + XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize; + XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX); + + { XXH128_hash_t acc; + int const nbRounds = (int)len / 32; + int i; + acc.low64 = len * PRIME64_1; + acc.high64 = 0; + for (i=0; i<4; i++) { + acc = XXH128_mix32B(acc, input+(32*i), input+(32*i)+16, secret+(32*i), seed); + } + acc.low64 = XXH3_avalanche(acc.low64); + acc.high64 = XXH3_avalanche(acc.high64); + XXH_ASSERT(nbRounds >= 4); + for (i=4 ; i < nbRounds; i++) { + acc = XXH128_mix32B(acc, input+(32*i), input+(32*i)+16, secret+XXH3_MIDSIZE_STARTOFFSET+(32*(i-4)), seed); + } + /* last bytes */ + acc = XXH128_mix32B(acc, input + len - 16, input + len - 32, secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET - 16, 0ULL - seed); + + { xxh_u64 const low64 = acc.low64 + acc.high64; + xxh_u64 const high64 = (acc.low64 * PRIME64_1) + (acc.high64 * PRIME64_4) + ((len - seed) * PRIME64_2); + XXH128_hash_t const h128 = { XXH3_avalanche(low64), (XXH64_hash_t)0 - XXH3_avalanche(high64) }; + return h128; + } + } +} + + +XXH_FORCE_INLINE XXH128_hash_t +XXH3_len_17to128_128b(const xxh_u8* XXH_RESTRICT input, size_t len, + const xxh_u8* XXH_RESTRICT secret, size_t secretSize, + XXH64_hash_t seed) +{ + XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize; + XXH_ASSERT(16 < len && len <= 128); + + { XXH128_hash_t acc; + acc.low64 = len * PRIME64_1; + acc.high64 = 0; + if (len > 32) { + if (len > 64) { + if (len > 96) { + acc = XXH128_mix32B(acc, input+48, input+len-64, secret+96, seed); + } + acc = XXH128_mix32B(acc, input+32, input+len-48, secret+64, seed); + } + acc = XXH128_mix32B(acc, input+16, input+len-32, secret+32, seed); + } + acc = XXH128_mix32B(acc, input, input+len-16, secret, seed); + { xxh_u64 const low64 = acc.low64 + acc.high64; + xxh_u64 const high64 = (acc.low64 * PRIME64_1) + (acc.high64 * PRIME64_4) + ((len - seed) * PRIME64_2); + XXH128_hash_t const h128 = { XXH3_avalanche(low64), (XXH64_hash_t)0 - XXH3_avalanche(high64) }; + return h128; + } + } +} + +XXH_PUBLIC_API XXH128_hash_t XXH3_128bits(const void* input, size_t len) +{ + if (len <= 16) return XXH3_len_0to16_128b((const xxh_u8*)input, len, kSecret, 0); + if (len <= 128) return XXH3_len_17to128_128b((const xxh_u8*)input, len, kSecret, sizeof(kSecret), 0); + if (len <= XXH3_MIDSIZE_MAX) return XXH3_len_129to240_128b((const xxh_u8*)input, len, kSecret, sizeof(kSecret), 0); + return XXH3_hashLong_128b_defaultSecret((const xxh_u8*)input, len); +} + +XXH_PUBLIC_API XXH128_hash_t +XXH3_128bits_withSecret(const void* input, size_t len, const void* secret, size_t secretSize) +{ + XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); + /* if an action must be taken should `secret` conditions not be respected, + * it should be done here. + * For now, it's a contract pre-condition. + * Adding a check and a branch here would cost performance at every hash */ + if (len <= 16) return XXH3_len_0to16_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, 0); + if (len <= 128) return XXH3_len_17to128_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretSize, 0); + if (len <= XXH3_MIDSIZE_MAX) return XXH3_len_129to240_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretSize, 0); + return XXH3_hashLong_128b_withSecret((const xxh_u8*)input, len, (const xxh_u8*)secret, secretSize); +} + +XXH_PUBLIC_API XXH128_hash_t +XXH3_128bits_withSeed(const void* input, size_t len, XXH64_hash_t seed) +{ + if (len <= 16) return XXH3_len_0to16_128b((const xxh_u8*)input, len, kSecret, seed); + if (len <= 128) return XXH3_len_17to128_128b((const xxh_u8*)input, len, kSecret, sizeof(kSecret), seed); + if (len <= XXH3_MIDSIZE_MAX) return XXH3_len_129to240_128b((const xxh_u8*)input, len, kSecret, sizeof(kSecret), seed); + return XXH3_hashLong_128b_withSeed((const xxh_u8*)input, len, seed); +} + +XXH_PUBLIC_API XXH128_hash_t +XXH128(const void* input, size_t len, XXH64_hash_t seed) +{ + return XXH3_128bits_withSeed(input, len, seed); +} + + +/* === XXH3 128-bit streaming === */ + +/* all the functions are actually the same as for 64-bit streaming variant, + just the reset one is different (different initial acc values for 0,5,6,7), + and near the end of the digest function */ + +static void +XXH3_128bits_reset_internal(XXH3_state_t* statePtr, + XXH64_hash_t seed, + const xxh_u8* secret, size_t secretSize) +{ + XXH3_64bits_reset_internal(statePtr, seed, secret, secretSize); +} + +XXH_PUBLIC_API XXH_errorcode +XXH3_128bits_reset(XXH3_state_t* statePtr) +{ + if (statePtr == NULL) return XXH_ERROR; + XXH3_128bits_reset_internal(statePtr, 0, kSecret, XXH_SECRET_DEFAULT_SIZE); + return XXH_OK; +} + +XXH_PUBLIC_API XXH_errorcode +XXH3_128bits_reset_withSecret(XXH3_state_t* statePtr, const void* secret, size_t secretSize) +{ + if (statePtr == NULL) return XXH_ERROR; + XXH3_128bits_reset_internal(statePtr, 0, (const xxh_u8*)secret, secretSize); + if (secret == NULL) return XXH_ERROR; + if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR; + return XXH_OK; +} + +XXH_PUBLIC_API XXH_errorcode +XXH3_128bits_reset_withSeed(XXH3_state_t* statePtr, XXH64_hash_t seed) +{ + if (statePtr == NULL) return XXH_ERROR; + XXH3_128bits_reset_internal(statePtr, seed, kSecret, XXH_SECRET_DEFAULT_SIZE); + XXH3_initCustomSecret(statePtr->customSecret, seed); + statePtr->secret = statePtr->customSecret; + return XXH_OK; +} + +XXH_PUBLIC_API XXH_errorcode +XXH3_128bits_update(XXH3_state_t* state, const void* input, size_t len) +{ + return XXH3_update(state, (const xxh_u8*)input, len, XXH3_acc_128bits); +} + +XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_digest (const XXH3_state_t* state) +{ + if (state->totalLen > XXH3_MIDSIZE_MAX) { + XXH_ALIGN(XXH_ACC_ALIGN) XXH64_hash_t acc[ACC_NB]; + XXH3_digest_long(acc, state, XXH3_acc_128bits); + XXH_ASSERT(state->secretLimit + STRIPE_LEN >= sizeof(acc) + XXH_SECRET_MERGEACCS_START); + { xxh_u64 const low64 = XXH3_mergeAccs(acc, state->secret + XXH_SECRET_MERGEACCS_START, (xxh_u64)state->totalLen * PRIME64_1); + xxh_u64 const high64 = XXH3_mergeAccs(acc, state->secret + state->secretLimit + STRIPE_LEN - sizeof(acc) - XXH_SECRET_MERGEACCS_START, ~((xxh_u64)state->totalLen * PRIME64_2)); + XXH128_hash_t const h128 = { low64, high64 }; + return h128; + } + } + /* len <= XXH3_MIDSIZE_MAX : short code */ + if (state->seed) + return XXH3_128bits_withSeed(state->buffer, (size_t)state->totalLen, state->seed); + return XXH3_128bits_withSecret(state->buffer, (size_t)(state->totalLen), state->secret, state->secretLimit + STRIPE_LEN); +} + +/* 128-bit utility functions */ + +#include /* memcmp */ + +/* return : 1 is equal, 0 if different */ +XXH_PUBLIC_API int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2) +{ + /* note : XXH128_hash_t is compact, it has no padding byte */ + return !(memcmp(&h1, &h2, sizeof(h1))); +} + +/* This prototype is compatible with stdlib's qsort(). + * return : >0 if *h128_1 > *h128_2 + * <0 if *h128_1 < *h128_2 + * =0 if *h128_1 == *h128_2 */ +XXH_PUBLIC_API int XXH128_cmp(const void* h128_1, const void* h128_2) +{ + XXH128_hash_t const h1 = *(const XXH128_hash_t*)h128_1; + XXH128_hash_t const h2 = *(const XXH128_hash_t*)h128_2; + int const hcmp = (h1.high64 > h2.high64) - (h2.high64 > h1.high64); + /* note : bets that, in most cases, hash values are different */ + if (hcmp) return hcmp; + return (h1.low64 > h2.low64) - (h2.low64 > h1.low64); +} + + +/*====== Canonical representation ======*/ +XXH_PUBLIC_API void +XXH128_canonicalFromHash(XXH128_canonical_t* dst, XXH128_hash_t hash) +{ + XXH_STATIC_ASSERT(sizeof(XXH128_canonical_t) == sizeof(XXH128_hash_t)); + if (XXH_CPU_LITTLE_ENDIAN) { + hash.high64 = XXH_swap64(hash.high64); + hash.low64 = XXH_swap64(hash.low64); + } + memcpy(dst, &hash.high64, sizeof(hash.high64)); + memcpy((char*)dst + sizeof(hash.high64), &hash.low64, sizeof(hash.low64)); +} + +XXH_PUBLIC_API XXH128_hash_t +XXH128_hashFromCanonical(const XXH128_canonical_t* src) +{ + XXH128_hash_t h; + h.high64 = XXH_readBE64(src); + h.low64 = XXH_readBE64(src->digest + 8); + return h; +} + + + +#endif /* XXH3_H */ diff --git a/src/xxHash/xxhash.c b/src/xxHash/xxhash.c index da06ea72..3f49b7d1 100644 --- a/src/xxHash/xxhash.c +++ b/src/xxHash/xxhash.c @@ -33,6 +33,12 @@ */ +/* since xxhash.c can be included (via XXH_INLINE_ALL), + * it's good practice to protect it with guard + * in case of multiples inclusions */ +#ifndef XXHASH_C_01393879 +#define XXHASH_C_01393879 + /* ************************************* * Tuning parameters ***************************************/ @@ -50,14 +56,10 @@ * Prefer these methods in priority order (0 > 1 > 2) */ #ifndef XXH_FORCE_MEMORY_ACCESS /* can be defined externally, on command line for example */ -# if defined(__GNUC__) && ( defined(__ARM_ARCH_6__) || defined(__ARM_ARCH_6J__) \ - || defined(__ARM_ARCH_6K__) || defined(__ARM_ARCH_6Z__) \ - || defined(__ARM_ARCH_6ZK__) || defined(__ARM_ARCH_6T2__) ) +# if !defined(__clang__) && defined(__GNUC__) && defined(__ARM_FEATURE_UNALIGNED) && defined(__ARM_ARCH) && (__ARM_ARCH == 6) # define XXH_FORCE_MEMORY_ACCESS 2 -# elif (defined(__INTEL_COMPILER) && !defined(_WIN32)) || \ - (defined(__GNUC__) && ( defined(__ARM_ARCH_7__) || defined(__ARM_ARCH_7A__) \ - || defined(__ARM_ARCH_7R__) || defined(__ARM_ARCH_7M__) \ - || defined(__ARM_ARCH_7S__) )) +# elif !defined(__clang__) && ((defined(__INTEL_COMPILER) && !defined(_WIN32)) || \ + (defined(__GNUC__) && (defined(__ARM_ARCH) && __ARM_ARCH >= 7))) # define XXH_FORCE_MEMORY_ACCESS 1 # endif #endif @@ -71,18 +73,6 @@ # define XXH_ACCEPT_NULL_INPUT_POINTER 0 #endif -/*!XXH_FORCE_NATIVE_FORMAT : - * By default, xxHash library provides endian-independent Hash values, based on little-endian convention. - * Results are therefore identical for little-endian and big-endian CPU. - * This comes at a performance cost for big-endian CPU, since some swapping is required to emulate little-endian format. - * Should endian-independence be of no importance for your application, you may set the #define below to 1, - * to improve speed for Big-endian CPU. - * This option has no impact on Little_Endian CPU. - */ -#ifndef XXH_FORCE_NATIVE_FORMAT /* can be defined externally */ -# define XXH_FORCE_NATIVE_FORMAT 0 -#endif - /*!XXH_FORCE_ALIGN_CHECK : * This is a minor performance trick, only useful with lots of very small keys. * It means : check for aligned/unaligned input. @@ -98,6 +88,18 @@ # endif #endif +/*!XXH_REROLL: + * Whether to reroll XXH32_finalize, and XXH64_finalize, + * instead of using an unrolled jump table/if statement loop. + * + * This is automatically defined on -Os/-Oz on GCC and Clang. */ +#ifndef XXH_REROLL +# if defined(__OPTIMIZE_SIZE__) +# define XXH_REROLL 1 +# else +# define XXH_REROLL 0 +# endif +#endif /* ************************************* * Includes & Memory related functions @@ -111,7 +113,7 @@ static void XXH_free (void* p) { free(p); } #include static void* XXH_memcpy(void* dest, const void* src, size_t size) { return memcpy(dest,src,size); } -#include /* assert */ +#include /* ULLONG_MAX */ #define XXH_STATIC_LINKING_ONLY #include "xxhash.h" @@ -122,58 +124,82 @@ static void* XXH_memcpy(void* dest, const void* src, size_t size) { return memcp ***************************************/ #ifdef _MSC_VER /* Visual Studio */ # pragma warning(disable : 4127) /* disable: C4127: conditional expression is constant */ -# define FORCE_INLINE static __forceinline +# define XXH_FORCE_INLINE static __forceinline +# define XXH_NO_INLINE static __declspec(noinline) #else # if defined (__cplusplus) || defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L /* C99 */ # ifdef __GNUC__ -# define FORCE_INLINE static inline __attribute__((always_inline)) +# define XXH_FORCE_INLINE static inline __attribute__((always_inline)) +# define XXH_NO_INLINE static __attribute__((noinline)) # else -# define FORCE_INLINE static inline +# define XXH_FORCE_INLINE static inline +# define XXH_NO_INLINE static # endif # else -# define FORCE_INLINE static +# define XXH_FORCE_INLINE static +# define XXH_NO_INLINE static # endif /* __STDC_VERSION__ */ #endif + +/* ************************************* +* Debug +***************************************/ +/* DEBUGLEVEL is expected to be defined externally, + * typically through compiler command line. + * Value must be a number. */ +#ifndef DEBUGLEVEL +# define DEBUGLEVEL 0 +#endif + +#if (DEBUGLEVEL>=1) +# include /* note : can still be disabled with NDEBUG */ +# define XXH_ASSERT(c) assert(c) +#else +# define XXH_ASSERT(c) ((void)0) +#endif + +/* note : use after variable declarations */ +#define XXH_STATIC_ASSERT(c) { enum { XXH_sa = 1/(int)(!!(c)) }; } + + /* ************************************* * Basic Types ***************************************/ -#ifndef MEM_MODULE -# if !defined (__VMS) \ - && (defined (__cplusplus) \ - || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) ) -# include - typedef uint8_t BYTE; - typedef uint16_t U16; - typedef uint32_t U32; -# else - typedef unsigned char BYTE; - typedef unsigned short U16; - typedef unsigned int U32; -# endif +#if !defined (__VMS) \ + && (defined (__cplusplus) \ + || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) ) +# include + typedef uint8_t xxh_u8; +#else + typedef unsigned char xxh_u8; #endif +typedef XXH32_hash_t xxh_u32; + + +/* === Memory access === */ #if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==2)) /* Force direct memory access. Only works on CPU which support unaligned memory access in hardware */ -static U32 XXH_read32(const void* memPtr) { return *(const U32*) memPtr; } +static xxh_u32 XXH_read32(const void* memPtr) { return *(const xxh_u32*) memPtr; } #elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==1)) /* __pack instructions are safer, but compiler specific, hence potentially problematic for some compilers */ /* currently only defined for gcc and icc */ -typedef union { U32 u32; } __attribute__((packed)) unalign; -static U32 XXH_read32(const void* ptr) { return ((const unalign*)ptr)->u32; } +typedef union { xxh_u32 u32; } __attribute__((packed)) unalign; +static xxh_u32 XXH_read32(const void* ptr) { return ((const unalign*)ptr)->u32; } #else /* portable and safe solution. Generally efficient. * see : http://stackoverflow.com/a/32095106/646947 */ -static U32 XXH_read32(const void* memPtr) +static xxh_u32 XXH_read32(const void* memPtr) { - U32 val; + xxh_u32 val; memcpy(&val, memPtr, sizeof(val)); return val; } @@ -181,18 +207,50 @@ static U32 XXH_read32(const void* memPtr) #endif /* XXH_FORCE_DIRECT_MEMORY_ACCESS */ +/* === Endianess === */ +typedef enum { XXH_bigEndian=0, XXH_littleEndian=1 } XXH_endianess; + +/* XXH_CPU_LITTLE_ENDIAN can be defined externally, for example on the compiler command line */ +#ifndef XXH_CPU_LITTLE_ENDIAN +# if defined(_WIN32) /* Windows is always little endian */ \ + || defined(__LITTLE_ENDIAN__) \ + || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__) +# define XXH_CPU_LITTLE_ENDIAN 1 +# elif defined(__BIG_ENDIAN__) \ + || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__) +# define XXH_CPU_LITTLE_ENDIAN 0 +# else +static int XXH_isLittleEndian(void) +{ + const union { xxh_u32 u; xxh_u8 c[4]; } one = { 1 }; /* don't use static : performance detrimental */ + return one.c[0]; +} +# define XXH_CPU_LITTLE_ENDIAN XXH_isLittleEndian() +# endif +#endif + + + + /* **************************************** * Compiler-specific Functions and Macros ******************************************/ #define XXH_GCC_VERSION (__GNUC__ * 100 + __GNUC_MINOR__) +#ifndef __has_builtin +# define __has_builtin(x) 0 +#endif + +#if !defined(NO_CLANG_BUILTIN) && __has_builtin(__builtin_rotateleft32) && __has_builtin(__builtin_rotateleft64) +# define XXH_rotl32 __builtin_rotateleft32 +# define XXH_rotl64 __builtin_rotateleft64 /* Note : although _rotl exists for minGW (GCC under windows), performance seems poor */ -#if defined(_MSC_VER) +#elif defined(_MSC_VER) # define XXH_rotl32(x,r) _rotl(x,r) # define XXH_rotl64(x,r) _rotl64(x,r) #else -# define XXH_rotl32(x,r) ((x << r) | (x >> (32 - r))) -# define XXH_rotl64(x,r) ((x << r) | (x >> (64 - r))) +# define XXH_rotl32(x,r) (((x) << (r)) | ((x) >> (32 - (r)))) +# define XXH_rotl64(x,r) (((x) << (r)) | ((x) >> (64 - (r)))) #endif #if defined(_MSC_VER) /* Visual Studio */ @@ -200,7 +258,7 @@ static U32 XXH_read32(const void* memPtr) #elif XXH_GCC_VERSION >= 403 # define XXH_swap32 __builtin_bswap32 #else -static U32 XXH_swap32 (U32 x) +static xxh_u32 XXH_swap32 (xxh_u32 x) { return ((x << 24) & 0xff000000 ) | ((x << 8) & 0x00ff0000 ) | @@ -210,72 +268,101 @@ static U32 XXH_swap32 (U32 x) #endif -/* ************************************* -* Architecture Macros -***************************************/ -typedef enum { XXH_bigEndian=0, XXH_littleEndian=1 } XXH_endianess; - -/* XXH_CPU_LITTLE_ENDIAN can be defined externally, for example on the compiler command line */ -#ifndef XXH_CPU_LITTLE_ENDIAN -static int XXH_isLittleEndian(void) -{ - const union { U32 u; BYTE c[4]; } one = { 1 }; /* don't use static : performance detrimental */ - return one.c[0]; -} -# define XXH_CPU_LITTLE_ENDIAN XXH_isLittleEndian() -#endif - - /* *************************** * Memory reads *****************************/ typedef enum { XXH_aligned, XXH_unaligned } XXH_alignment; -FORCE_INLINE U32 XXH_readLE32_align(const void* ptr, XXH_endianess endian, XXH_alignment align) +XXH_FORCE_INLINE xxh_u32 XXH_readLE32(const void* ptr) { - if (align==XXH_unaligned) - return endian==XXH_littleEndian ? XXH_read32(ptr) : XXH_swap32(XXH_read32(ptr)); - else - return endian==XXH_littleEndian ? *(const U32*)ptr : XXH_swap32(*(const U32*)ptr); + return XXH_CPU_LITTLE_ENDIAN ? XXH_read32(ptr) : XXH_swap32(XXH_read32(ptr)); } -FORCE_INLINE U32 XXH_readLE32(const void* ptr, XXH_endianess endian) -{ - return XXH_readLE32_align(ptr, endian, XXH_unaligned); -} - -static U32 XXH_readBE32(const void* ptr) +static xxh_u32 XXH_readBE32(const void* ptr) { return XXH_CPU_LITTLE_ENDIAN ? XXH_swap32(XXH_read32(ptr)) : XXH_read32(ptr); } +XXH_FORCE_INLINE xxh_u32 +XXH_readLE32_align(const void* ptr, XXH_alignment align) +{ + if (align==XXH_unaligned) { + return XXH_readLE32(ptr); + } else { + return XXH_CPU_LITTLE_ENDIAN ? *(const xxh_u32*)ptr : XXH_swap32(*(const xxh_u32*)ptr); + } +} + /* ************************************* -* Macros +* Misc ***************************************/ -#define XXH_STATIC_ASSERT(c) { enum { XXH_sa = 1/(int)(!!(c)) }; } /* use after variable declarations */ XXH_PUBLIC_API unsigned XXH_versionNumber (void) { return XXH_VERSION_NUMBER; } /* ******************************************************************* * 32-bit hash functions *********************************************************************/ -static const U32 PRIME32_1 = 2654435761U; -static const U32 PRIME32_2 = 2246822519U; -static const U32 PRIME32_3 = 3266489917U; -static const U32 PRIME32_4 = 668265263U; -static const U32 PRIME32_5 = 374761393U; +static const xxh_u32 PRIME32_1 = 0x9E3779B1U; /* 0b10011110001101110111100110110001 */ +static const xxh_u32 PRIME32_2 = 0x85EBCA77U; /* 0b10000101111010111100101001110111 */ +static const xxh_u32 PRIME32_3 = 0xC2B2AE3DU; /* 0b11000010101100101010111000111101 */ +static const xxh_u32 PRIME32_4 = 0x27D4EB2FU; /* 0b00100111110101001110101100101111 */ +static const xxh_u32 PRIME32_5 = 0x165667B1U; /* 0b00010110010101100110011110110001 */ -static U32 XXH32_round(U32 seed, U32 input) +static xxh_u32 XXH32_round(xxh_u32 acc, xxh_u32 input) { - seed += input * PRIME32_2; - seed = XXH_rotl32(seed, 13); - seed *= PRIME32_1; - return seed; + acc += input * PRIME32_2; + acc = XXH_rotl32(acc, 13); + acc *= PRIME32_1; +#if defined(__GNUC__) && defined(__SSE4_1__) && !defined(XXH_ENABLE_AUTOVECTORIZE) + /* UGLY HACK: + * This inline assembly hack forces acc into a normal register. This is the + * only thing that prevents GCC and Clang from autovectorizing the XXH32 loop + * (pragmas and attributes don't work for some resason) without globally + * disabling SSE4.1. + * + * The reason we want to avoid vectorization is because despite working on + * 4 integers at a time, there are multiple factors slowing XXH32 down on + * SSE4: + * - There's a ridiculous amount of lag from pmulld (10 cycles of latency on newer chips!) + * making it slightly slower to multiply four integers at once compared to four + * integers independently. Even when pmulld was fastest, Sandy/Ivy Bridge, it is + * still not worth it to go into SSE just to multiply unless doing a long operation. + * + * - Four instructions are required to rotate, + * movqda tmp, v // not required with VEX encoding + * pslld tmp, 13 // tmp <<= 13 + * psrld v, 19 // x >>= 19 + * por v, tmp // x |= tmp + * compared to one for scalar: + * roll v, 13 // reliably fast across the board + * shldl v, v, 13 // Sandy Bridge and later prefer this for some reason + * + * - Instruction level parallelism is actually more beneficial here because the + * SIMD actually serializes this operation: While v1 is rotating, v2 can load data, + * while v3 can multiply. SSE forces them to operate together. + * + * How this hack works: + * __asm__("" // Declare an assembly block but don't declare any instructions + * : // However, as an Input/Output Operand, + * "+r" // constrain a read/write operand (+) as a general purpose register (r). + * (acc) // and set acc as the operand + * ); + * + * Because of the 'r', the compiler has promised that seed will be in a + * general purpose register and the '+' says that it will be 'read/write', + * so it has to assume it has changed. It is like volatile without all the + * loads and stores. + * + * Since the argument has to be in a normal register (not an SSE register), + * each time XXH32_round is called, it is impossible to vectorize. */ + __asm__("" : "+r" (acc)); +#endif + return acc; } /* mix all bits */ -static U32 XXH32_avalanche(U32 h32) +static xxh_u32 XXH32_avalanche(xxh_u32 h32) { h32 ^= h32 >> 15; h32 *= PRIME32_2; @@ -285,97 +372,103 @@ static U32 XXH32_avalanche(U32 h32) return(h32); } -#define XXH_get32bits(p) XXH_readLE32_align(p, endian, align) - -static U32 -XXH32_finalize(U32 h32, const void* ptr, size_t len, - XXH_endianess endian, XXH_alignment align) +#define XXH_get32bits(p) XXH_readLE32_align(p, align) +static xxh_u32 +XXH32_finalize(xxh_u32 h32, const xxh_u8* ptr, size_t len, XXH_alignment align) { - const BYTE* p = (const BYTE*)ptr; -#define PROCESS1 \ - h32 += (*p) * PRIME32_5; \ - p++; \ +#define PROCESS1 \ + h32 += (*ptr++) * PRIME32_5; \ h32 = XXH_rotl32(h32, 11) * PRIME32_1 ; #define PROCESS4 \ - h32 += XXH_get32bits(p) * PRIME32_3; \ - p+=4; \ + h32 += XXH_get32bits(ptr) * PRIME32_3; \ + ptr+=4; \ h32 = XXH_rotl32(h32, 17) * PRIME32_4 ; - switch(len&15) /* or switch(bEnd - p) */ - { - case 12: PROCESS4; - /* fallthrough */ - case 8: PROCESS4; - /* fallthrough */ - case 4: PROCESS4; - return XXH32_avalanche(h32); + /* Compact rerolled version */ + if (XXH_REROLL) { + len &= 15; + while (len >= 4) { + PROCESS4; + len -= 4; + } + while (len > 0) { + PROCESS1; + --len; + } + return XXH32_avalanche(h32); + } else { + switch(len&15) /* or switch(bEnd - p) */ { + case 12: PROCESS4; + /* fallthrough */ + case 8: PROCESS4; + /* fallthrough */ + case 4: PROCESS4; + return XXH32_avalanche(h32); - case 13: PROCESS4; - /* fallthrough */ - case 9: PROCESS4; - /* fallthrough */ - case 5: PROCESS4; - PROCESS1; - return XXH32_avalanche(h32); + case 13: PROCESS4; + /* fallthrough */ + case 9: PROCESS4; + /* fallthrough */ + case 5: PROCESS4; + PROCESS1; + return XXH32_avalanche(h32); - case 14: PROCESS4; - /* fallthrough */ - case 10: PROCESS4; - /* fallthrough */ - case 6: PROCESS4; - PROCESS1; - PROCESS1; - return XXH32_avalanche(h32); + case 14: PROCESS4; + /* fallthrough */ + case 10: PROCESS4; + /* fallthrough */ + case 6: PROCESS4; + PROCESS1; + PROCESS1; + return XXH32_avalanche(h32); - case 15: PROCESS4; - /* fallthrough */ - case 11: PROCESS4; - /* fallthrough */ - case 7: PROCESS4; - /* fallthrough */ - case 3: PROCESS1; - /* fallthrough */ - case 2: PROCESS1; - /* fallthrough */ - case 1: PROCESS1; - /* fallthrough */ - case 0: return XXH32_avalanche(h32); + case 15: PROCESS4; + /* fallthrough */ + case 11: PROCESS4; + /* fallthrough */ + case 7: PROCESS4; + /* fallthrough */ + case 3: PROCESS1; + /* fallthrough */ + case 2: PROCESS1; + /* fallthrough */ + case 1: PROCESS1; + /* fallthrough */ + case 0: return XXH32_avalanche(h32); + } + XXH_ASSERT(0); + return h32; /* reaching this point is deemed impossible */ } - assert(0); - return h32; /* reaching this point is deemed impossible */ } - -FORCE_INLINE U32 -XXH32_endian_align(const void* input, size_t len, U32 seed, - XXH_endianess endian, XXH_alignment align) +XXH_FORCE_INLINE xxh_u32 +XXH32_endian_align(const xxh_u8* input, size_t len, xxh_u32 seed, XXH_alignment align) { - const BYTE* p = (const BYTE*)input; - const BYTE* bEnd = p + len; - U32 h32; + const xxh_u8* bEnd = input + len; + xxh_u32 h32; #if defined(XXH_ACCEPT_NULL_INPUT_POINTER) && (XXH_ACCEPT_NULL_INPUT_POINTER>=1) - if (p==NULL) { + if (input==NULL) { len=0; - bEnd=p=(const BYTE*)(size_t)16; + bEnd=input=(const xxh_u8*)(size_t)16; } #endif if (len>=16) { - const BYTE* const limit = bEnd - 15; - U32 v1 = seed + PRIME32_1 + PRIME32_2; - U32 v2 = seed + PRIME32_2; - U32 v3 = seed + 0; - U32 v4 = seed - PRIME32_1; + const xxh_u8* const limit = bEnd - 15; + xxh_u32 v1 = seed + PRIME32_1 + PRIME32_2; + xxh_u32 v2 = seed + PRIME32_2; + xxh_u32 v3 = seed + 0; + xxh_u32 v4 = seed - PRIME32_1; do { - v1 = XXH32_round(v1, XXH_get32bits(p)); p+=4; - v2 = XXH32_round(v2, XXH_get32bits(p)); p+=4; - v3 = XXH32_round(v3, XXH_get32bits(p)); p+=4; - v4 = XXH32_round(v4, XXH_get32bits(p)); p+=4; - } while (p < limit); + v1 = XXH32_round(v1, XXH_get32bits(input)); input += 4; + v2 = XXH32_round(v2, XXH_get32bits(input)); input += 4; + v3 = XXH32_round(v3, XXH_get32bits(input)); input += 4; + v4 = XXH32_round(v4, XXH_get32bits(input)); input += 4; + } while (input < limit); h32 = XXH_rotl32(v1, 1) + XXH_rotl32(v2, 7) + XXH_rotl32(v3, 12) + XXH_rotl32(v4, 18); @@ -383,35 +476,29 @@ XXH32_endian_align(const void* input, size_t len, U32 seed, h32 = seed + PRIME32_5; } - h32 += (U32)len; + h32 += (xxh_u32)len; - return XXH32_finalize(h32, p, len&15, endian, align); + return XXH32_finalize(h32, input, len&15, align); } -XXH_PUBLIC_API unsigned int XXH32 (const void* input, size_t len, unsigned int seed) +XXH_PUBLIC_API XXH32_hash_t XXH32 (const void* input, size_t len, XXH32_hash_t seed) { #if 0 /* Simple version, good for code maintenance, but unfortunately slow for small inputs */ XXH32_state_t state; XXH32_reset(&state, seed); - XXH32_update(&state, input, len); + XXH32_update(&state, (const xxh_u8*)input, len); return XXH32_digest(&state); + #else - XXH_endianess endian_detected = (XXH_endianess)XXH_CPU_LITTLE_ENDIAN; if (XXH_FORCE_ALIGN_CHECK) { if ((((size_t)input) & 3) == 0) { /* Input is 4-bytes aligned, leverage the speed benefit */ - if ((endian_detected==XXH_littleEndian) || XXH_FORCE_NATIVE_FORMAT) - return XXH32_endian_align(input, len, seed, XXH_littleEndian, XXH_aligned); - else - return XXH32_endian_align(input, len, seed, XXH_bigEndian, XXH_aligned); + return XXH32_endian_align((const xxh_u8*)input, len, seed, XXH_aligned); } } - if ((endian_detected==XXH_littleEndian) || XXH_FORCE_NATIVE_FORMAT) - return XXH32_endian_align(input, len, seed, XXH_littleEndian, XXH_unaligned); - else - return XXH32_endian_align(input, len, seed, XXH_bigEndian, XXH_unaligned); + return XXH32_endian_align((const xxh_u8*)input, len, seed, XXH_unaligned); #endif } @@ -434,7 +521,7 @@ XXH_PUBLIC_API void XXH32_copyState(XXH32_state_t* dstState, const XXH32_state_t memcpy(dstState, srcState, sizeof(*dstState)); } -XXH_PUBLIC_API XXH_errorcode XXH32_reset(XXH32_state_t* statePtr, unsigned int seed) +XXH_PUBLIC_API XXH_errorcode XXH32_reset(XXH32_state_t* statePtr, XXH32_hash_t seed) { XXH32_state_t state; /* using a local state to memcpy() in order to avoid strict-aliasing warnings */ memset(&state, 0, sizeof(state)); @@ -448,12 +535,9 @@ XXH_PUBLIC_API XXH_errorcode XXH32_reset(XXH32_state_t* statePtr, unsigned int s } -FORCE_INLINE -XXH_errorcode XXH32_update_endian (XXH32_state_t* state, const void* input, size_t len, XXH_endianess endian) +XXH_PUBLIC_API XXH_errorcode +XXH32_update(XXH32_state_t* state, const void* input, size_t len) { - const BYTE* p = (const BYTE*)input; - const BYTE* const bEnd = p + len; - if (input==NULL) #if defined(XXH_ACCEPT_NULL_INPUT_POINTER) && (XXH_ACCEPT_NULL_INPUT_POINTER>=1) return XXH_OK; @@ -461,71 +545,63 @@ XXH_errorcode XXH32_update_endian (XXH32_state_t* state, const void* input, size return XXH_ERROR; #endif - state->total_len_32 += (unsigned)len; - state->large_len |= (len>=16) | (state->total_len_32>=16); + { const xxh_u8* p = (const xxh_u8*)input; + const xxh_u8* const bEnd = p + len; - if (state->memsize + len < 16) { /* fill in tmp buffer */ - XXH_memcpy((BYTE*)(state->mem32) + state->memsize, input, len); - state->memsize += (unsigned)len; - return XXH_OK; - } + state->total_len_32 += (XXH32_hash_t)len; + state->large_len |= (XXH32_hash_t)((len>=16) | (state->total_len_32>=16)); - if (state->memsize) { /* some data left from previous update */ - XXH_memcpy((BYTE*)(state->mem32) + state->memsize, input, 16-state->memsize); - { const U32* p32 = state->mem32; - state->v1 = XXH32_round(state->v1, XXH_readLE32(p32, endian)); p32++; - state->v2 = XXH32_round(state->v2, XXH_readLE32(p32, endian)); p32++; - state->v3 = XXH32_round(state->v3, XXH_readLE32(p32, endian)); p32++; - state->v4 = XXH32_round(state->v4, XXH_readLE32(p32, endian)); + if (state->memsize + len < 16) { /* fill in tmp buffer */ + XXH_memcpy((xxh_u8*)(state->mem32) + state->memsize, input, len); + state->memsize += (XXH32_hash_t)len; + return XXH_OK; } - p += 16-state->memsize; - state->memsize = 0; - } - if (p <= bEnd-16) { - const BYTE* const limit = bEnd - 16; - U32 v1 = state->v1; - U32 v2 = state->v2; - U32 v3 = state->v3; - U32 v4 = state->v4; + if (state->memsize) { /* some data left from previous update */ + XXH_memcpy((xxh_u8*)(state->mem32) + state->memsize, input, 16-state->memsize); + { const xxh_u32* p32 = state->mem32; + state->v1 = XXH32_round(state->v1, XXH_readLE32(p32)); p32++; + state->v2 = XXH32_round(state->v2, XXH_readLE32(p32)); p32++; + state->v3 = XXH32_round(state->v3, XXH_readLE32(p32)); p32++; + state->v4 = XXH32_round(state->v4, XXH_readLE32(p32)); + } + p += 16-state->memsize; + state->memsize = 0; + } - do { - v1 = XXH32_round(v1, XXH_readLE32(p, endian)); p+=4; - v2 = XXH32_round(v2, XXH_readLE32(p, endian)); p+=4; - v3 = XXH32_round(v3, XXH_readLE32(p, endian)); p+=4; - v4 = XXH32_round(v4, XXH_readLE32(p, endian)); p+=4; - } while (p<=limit); + if (p <= bEnd-16) { + const xxh_u8* const limit = bEnd - 16; + xxh_u32 v1 = state->v1; + xxh_u32 v2 = state->v2; + xxh_u32 v3 = state->v3; + xxh_u32 v4 = state->v4; - state->v1 = v1; - state->v2 = v2; - state->v3 = v3; - state->v4 = v4; - } + do { + v1 = XXH32_round(v1, XXH_readLE32(p)); p+=4; + v2 = XXH32_round(v2, XXH_readLE32(p)); p+=4; + v3 = XXH32_round(v3, XXH_readLE32(p)); p+=4; + v4 = XXH32_round(v4, XXH_readLE32(p)); p+=4; + } while (p<=limit); - if (p < bEnd) { - XXH_memcpy(state->mem32, p, (size_t)(bEnd-p)); - state->memsize = (unsigned)(bEnd-p); + state->v1 = v1; + state->v2 = v2; + state->v3 = v3; + state->v4 = v4; + } + + if (p < bEnd) { + XXH_memcpy(state->mem32, p, (size_t)(bEnd-p)); + state->memsize = (unsigned)(bEnd-p); + } } return XXH_OK; } -XXH_PUBLIC_API XXH_errorcode XXH32_update (XXH32_state_t* state_in, const void* input, size_t len) +XXH_PUBLIC_API XXH32_hash_t XXH32_digest (const XXH32_state_t* state) { - XXH_endianess endian_detected = (XXH_endianess)XXH_CPU_LITTLE_ENDIAN; - - if ((endian_detected==XXH_littleEndian) || XXH_FORCE_NATIVE_FORMAT) - return XXH32_update_endian(state_in, input, len, XXH_littleEndian); - else - return XXH32_update_endian(state_in, input, len, XXH_bigEndian); -} - - -FORCE_INLINE U32 -XXH32_digest_endian (const XXH32_state_t* state, XXH_endianess endian) -{ - U32 h32; + xxh_u32 h32; if (state->large_len) { h32 = XXH_rotl32(state->v1, 1) @@ -538,18 +614,7 @@ XXH32_digest_endian (const XXH32_state_t* state, XXH_endianess endian) h32 += state->total_len_32; - return XXH32_finalize(h32, state->mem32, state->memsize, endian, XXH_aligned); -} - - -XXH_PUBLIC_API unsigned int XXH32_digest (const XXH32_state_t* state_in) -{ - XXH_endianess endian_detected = (XXH_endianess)XXH_CPU_LITTLE_ENDIAN; - - if ((endian_detected==XXH_littleEndian) || XXH_FORCE_NATIVE_FORMAT) - return XXH32_digest_endian(state_in, XXH_littleEndian); - else - return XXH32_digest_endian(state_in, XXH_bigEndian); + return XXH32_finalize(h32, (const xxh_u8*)state->mem32, state->memsize, XXH_aligned); } @@ -582,31 +647,46 @@ XXH_PUBLIC_API XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src /*====== Memory access ======*/ -#ifndef MEM_MODULE -# define MEM_MODULE -# if !defined (__VMS) \ - && (defined (__cplusplus) \ - || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) ) -# include - typedef uint64_t U64; -# else - /* if compiler doesn't support unsigned long long, replace by another 64-bit type */ - typedef unsigned long long U64; -# endif -#endif +typedef XXH64_hash_t xxh_u64; +/*! XXH_REROLL_XXH64: + * Whether to reroll the XXH64_finalize() loop. + * + * Just like XXH32, we can unroll the XXH64_finalize() loop. This can be a performance gain + * on 64-bit hosts, as only one jump is required. + * + * However, on 32-bit hosts, because arithmetic needs to be done with two 32-bit registers, + * and 64-bit arithmetic needs to be simulated, it isn't beneficial to unroll. The code becomes + * ridiculously large (the largest function in the binary on i386!), and rerolling it saves + * anywhere from 3kB to 20kB. It is also slightly faster because it fits into cache better + * and is more likely to be inlined by the compiler. + * + * If XXH_REROLL is defined, this is ignored and the loop is always rerolled. */ +#ifndef XXH_REROLL_XXH64 +# if (defined(__ILP32__) || defined(_ILP32)) /* ILP32 is often defined on 32-bit GCC family */ \ + || !(defined(__x86_64__) || defined(_M_X64) || defined(_M_AMD64) /* x86-64 */ \ + || defined(_M_ARM64) || defined(__aarch64__) || defined(__arm64__) /* aarch64 */ \ + || defined(__PPC64__) || defined(__PPC64LE__) || defined(__ppc64__) || defined(__powerpc64__) /* ppc64 */ \ + || defined(__mips64__) || defined(__mips64)) /* mips64 */ \ + || (!defined(SIZE_MAX) || SIZE_MAX < ULLONG_MAX) /* check limits */ +# define XXH_REROLL_XXH64 1 +# else +# define XXH_REROLL_XXH64 0 +# endif +#endif /* !defined(XXH_REROLL_XXH64) */ + #if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==2)) /* Force direct memory access. Only works on CPU which support unaligned memory access in hardware */ -static U64 XXH_read64(const void* memPtr) { return *(const U64*) memPtr; } +static xxh_u64 XXH_read64(const void* memPtr) { return *(const xxh_u64*) memPtr; } #elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==1)) /* __pack instructions are safer, but compiler specific, hence potentially problematic for some compilers */ /* currently only defined for gcc and icc */ -typedef union { U32 u32; U64 u64; } __attribute__((packed)) unalign64; -static U64 XXH_read64(const void* ptr) { return ((const unalign64*)ptr)->u64; } +typedef union { xxh_u32 u32; xxh_u64 u64; } __attribute__((packed)) unalign64; +static xxh_u64 XXH_read64(const void* ptr) { return ((const unalign64*)ptr)->u64; } #else @@ -614,9 +694,9 @@ static U64 XXH_read64(const void* ptr) { return ((const unalign64*)ptr)->u64; } * see : http://stackoverflow.com/a/32095106/646947 */ -static U64 XXH_read64(const void* memPtr) +static xxh_u64 XXH_read64(const void* memPtr) { - U64 val; + xxh_u64 val; memcpy(&val, memPtr, sizeof(val)); return val; } @@ -628,7 +708,7 @@ static U64 XXH_read64(const void* memPtr) #elif XXH_GCC_VERSION >= 403 # define XXH_swap64 __builtin_bswap64 #else -static U64 XXH_swap64 (U64 x) +static xxh_u64 XXH_swap64 (xxh_u64 x) { return ((x << 56) & 0xff00000000000000ULL) | ((x << 40) & 0x00ff000000000000ULL) | @@ -641,34 +721,35 @@ static U64 XXH_swap64 (U64 x) } #endif -FORCE_INLINE U64 XXH_readLE64_align(const void* ptr, XXH_endianess endian, XXH_alignment align) +XXH_FORCE_INLINE xxh_u64 XXH_readLE64(const void* ptr) { - if (align==XXH_unaligned) - return endian==XXH_littleEndian ? XXH_read64(ptr) : XXH_swap64(XXH_read64(ptr)); - else - return endian==XXH_littleEndian ? *(const U64*)ptr : XXH_swap64(*(const U64*)ptr); + return XXH_CPU_LITTLE_ENDIAN ? XXH_read64(ptr) : XXH_swap64(XXH_read64(ptr)); } -FORCE_INLINE U64 XXH_readLE64(const void* ptr, XXH_endianess endian) -{ - return XXH_readLE64_align(ptr, endian, XXH_unaligned); -} - -static U64 XXH_readBE64(const void* ptr) +static xxh_u64 XXH_readBE64(const void* ptr) { return XXH_CPU_LITTLE_ENDIAN ? XXH_swap64(XXH_read64(ptr)) : XXH_read64(ptr); } +XXH_FORCE_INLINE xxh_u64 +XXH_readLE64_align(const void* ptr, XXH_alignment align) +{ + if (align==XXH_unaligned) + return XXH_readLE64(ptr); + else + return XXH_CPU_LITTLE_ENDIAN ? *(const xxh_u64*)ptr : XXH_swap64(*(const xxh_u64*)ptr); +} + /*====== xxh64 ======*/ -static const U64 PRIME64_1 = 11400714785074694791ULL; -static const U64 PRIME64_2 = 14029467366897019727ULL; -static const U64 PRIME64_3 = 1609587929392839161ULL; -static const U64 PRIME64_4 = 9650029242287828579ULL; -static const U64 PRIME64_5 = 2870177450012600261ULL; +static const xxh_u64 PRIME64_1 = 0x9E3779B185EBCA87ULL; /* 0b1001111000110111011110011011000110000101111010111100101010000111 */ +static const xxh_u64 PRIME64_2 = 0xC2B2AE3D27D4EB4FULL; /* 0b1100001010110010101011100011110100100111110101001110101101001111 */ +static const xxh_u64 PRIME64_3 = 0x165667B19E3779F9ULL; /* 0b0001011001010110011001111011000110011110001101110111100111111001 */ +static const xxh_u64 PRIME64_4 = 0x85EBCA77C2B2AE63ULL; /* 0b1000010111101011110010100111011111000010101100101010111001100011 */ +static const xxh_u64 PRIME64_5 = 0x27D4EB2F165667C5ULL; /* 0b0010011111010100111010110010111100010110010101100110011111000101 */ -static U64 XXH64_round(U64 acc, U64 input) +static xxh_u64 XXH64_round(xxh_u64 acc, xxh_u64 input) { acc += input * PRIME64_2; acc = XXH_rotl64(acc, 31); @@ -676,7 +757,7 @@ static U64 XXH64_round(U64 acc, U64 input) return acc; } -static U64 XXH64_mergeRound(U64 acc, U64 val) +static xxh_u64 XXH64_mergeRound(xxh_u64 acc, xxh_u64 val) { val = XXH64_round(0, val); acc ^= val; @@ -684,7 +765,7 @@ static U64 XXH64_mergeRound(U64 acc, U64 val) return acc; } -static U64 XXH64_avalanche(U64 h64) +static xxh_u64 XXH64_avalanche(xxh_u64 h64) { h64 ^= h64 >> 33; h64 *= PRIME64_2; @@ -695,146 +776,157 @@ static U64 XXH64_avalanche(U64 h64) } -#define XXH_get64bits(p) XXH_readLE64_align(p, endian, align) +#define XXH_get64bits(p) XXH_readLE64_align(p, align) -static U64 -XXH64_finalize(U64 h64, const void* ptr, size_t len, - XXH_endianess endian, XXH_alignment align) +static xxh_u64 +XXH64_finalize(xxh_u64 h64, const xxh_u8* ptr, size_t len, XXH_alignment align) { - const BYTE* p = (const BYTE*)ptr; - -#define PROCESS1_64 \ - h64 ^= (*p) * PRIME64_5; \ - p++; \ +#define PROCESS1_64 \ + h64 ^= (*ptr++) * PRIME64_5; \ h64 = XXH_rotl64(h64, 11) * PRIME64_1; #define PROCESS4_64 \ - h64 ^= (U64)(XXH_get32bits(p)) * PRIME64_1; \ - p+=4; \ + h64 ^= (xxh_u64)(XXH_get32bits(ptr)) * PRIME64_1; \ + ptr+=4; \ h64 = XXH_rotl64(h64, 23) * PRIME64_2 + PRIME64_3; #define PROCESS8_64 { \ - U64 const k1 = XXH64_round(0, XXH_get64bits(p)); \ - p+=8; \ + xxh_u64 const k1 = XXH64_round(0, XXH_get64bits(ptr)); \ + ptr+=8; \ h64 ^= k1; \ h64 = XXH_rotl64(h64,27) * PRIME64_1 + PRIME64_4; \ } - switch(len&31) { - case 24: PROCESS8_64; - /* fallthrough */ - case 16: PROCESS8_64; - /* fallthrough */ - case 8: PROCESS8_64; - return XXH64_avalanche(h64); + /* Rerolled version for 32-bit targets is faster and much smaller. */ + if (XXH_REROLL || XXH_REROLL_XXH64) { + len &= 31; + while (len >= 8) { + PROCESS8_64; + len -= 8; + } + if (len >= 4) { + PROCESS4_64; + len -= 4; + } + while (len > 0) { + PROCESS1_64; + --len; + } + return XXH64_avalanche(h64); + } else { + switch(len & 31) { + case 24: PROCESS8_64; + /* fallthrough */ + case 16: PROCESS8_64; + /* fallthrough */ + case 8: PROCESS8_64; + return XXH64_avalanche(h64); - case 28: PROCESS8_64; - /* fallthrough */ - case 20: PROCESS8_64; - /* fallthrough */ - case 12: PROCESS8_64; - /* fallthrough */ - case 4: PROCESS4_64; - return XXH64_avalanche(h64); + case 28: PROCESS8_64; + /* fallthrough */ + case 20: PROCESS8_64; + /* fallthrough */ + case 12: PROCESS8_64; + /* fallthrough */ + case 4: PROCESS4_64; + return XXH64_avalanche(h64); - case 25: PROCESS8_64; - /* fallthrough */ - case 17: PROCESS8_64; - /* fallthrough */ - case 9: PROCESS8_64; - PROCESS1_64; - return XXH64_avalanche(h64); + case 25: PROCESS8_64; + /* fallthrough */ + case 17: PROCESS8_64; + /* fallthrough */ + case 9: PROCESS8_64; + PROCESS1_64; + return XXH64_avalanche(h64); - case 29: PROCESS8_64; - /* fallthrough */ - case 21: PROCESS8_64; - /* fallthrough */ - case 13: PROCESS8_64; - /* fallthrough */ - case 5: PROCESS4_64; - PROCESS1_64; - return XXH64_avalanche(h64); + case 29: PROCESS8_64; + /* fallthrough */ + case 21: PROCESS8_64; + /* fallthrough */ + case 13: PROCESS8_64; + /* fallthrough */ + case 5: PROCESS4_64; + PROCESS1_64; + return XXH64_avalanche(h64); - case 26: PROCESS8_64; - /* fallthrough */ - case 18: PROCESS8_64; - /* fallthrough */ - case 10: PROCESS8_64; - PROCESS1_64; - PROCESS1_64; - return XXH64_avalanche(h64); + case 26: PROCESS8_64; + /* fallthrough */ + case 18: PROCESS8_64; + /* fallthrough */ + case 10: PROCESS8_64; + PROCESS1_64; + PROCESS1_64; + return XXH64_avalanche(h64); - case 30: PROCESS8_64; - /* fallthrough */ - case 22: PROCESS8_64; - /* fallthrough */ - case 14: PROCESS8_64; - /* fallthrough */ - case 6: PROCESS4_64; - PROCESS1_64; - PROCESS1_64; - return XXH64_avalanche(h64); + case 30: PROCESS8_64; + /* fallthrough */ + case 22: PROCESS8_64; + /* fallthrough */ + case 14: PROCESS8_64; + /* fallthrough */ + case 6: PROCESS4_64; + PROCESS1_64; + PROCESS1_64; + return XXH64_avalanche(h64); - case 27: PROCESS8_64; - /* fallthrough */ - case 19: PROCESS8_64; - /* fallthrough */ - case 11: PROCESS8_64; - PROCESS1_64; - PROCESS1_64; - PROCESS1_64; - return XXH64_avalanche(h64); + case 27: PROCESS8_64; + /* fallthrough */ + case 19: PROCESS8_64; + /* fallthrough */ + case 11: PROCESS8_64; + PROCESS1_64; + PROCESS1_64; + PROCESS1_64; + return XXH64_avalanche(h64); - case 31: PROCESS8_64; - /* fallthrough */ - case 23: PROCESS8_64; - /* fallthrough */ - case 15: PROCESS8_64; - /* fallthrough */ - case 7: PROCESS4_64; - /* fallthrough */ - case 3: PROCESS1_64; - /* fallthrough */ - case 2: PROCESS1_64; - /* fallthrough */ - case 1: PROCESS1_64; - /* fallthrough */ - case 0: return XXH64_avalanche(h64); + case 31: PROCESS8_64; + /* fallthrough */ + case 23: PROCESS8_64; + /* fallthrough */ + case 15: PROCESS8_64; + /* fallthrough */ + case 7: PROCESS4_64; + /* fallthrough */ + case 3: PROCESS1_64; + /* fallthrough */ + case 2: PROCESS1_64; + /* fallthrough */ + case 1: PROCESS1_64; + /* fallthrough */ + case 0: return XXH64_avalanche(h64); + } } - /* impossible to reach */ - assert(0); + XXH_ASSERT(0); return 0; /* unreachable, but some compilers complain without it */ } -FORCE_INLINE U64 -XXH64_endian_align(const void* input, size_t len, U64 seed, - XXH_endianess endian, XXH_alignment align) +XXH_FORCE_INLINE xxh_u64 +XXH64_endian_align(const xxh_u8* input, size_t len, xxh_u64 seed, XXH_alignment align) { - const BYTE* p = (const BYTE*)input; - const BYTE* bEnd = p + len; - U64 h64; + const xxh_u8* bEnd = input + len; + xxh_u64 h64; #if defined(XXH_ACCEPT_NULL_INPUT_POINTER) && (XXH_ACCEPT_NULL_INPUT_POINTER>=1) - if (p==NULL) { + if (input==NULL) { len=0; - bEnd=p=(const BYTE*)(size_t)32; + bEnd=input=(const xxh_u8*)(size_t)32; } #endif if (len>=32) { - const BYTE* const limit = bEnd - 32; - U64 v1 = seed + PRIME64_1 + PRIME64_2; - U64 v2 = seed + PRIME64_2; - U64 v3 = seed + 0; - U64 v4 = seed - PRIME64_1; + const xxh_u8* const limit = bEnd - 32; + xxh_u64 v1 = seed + PRIME64_1 + PRIME64_2; + xxh_u64 v2 = seed + PRIME64_2; + xxh_u64 v3 = seed + 0; + xxh_u64 v4 = seed - PRIME64_1; do { - v1 = XXH64_round(v1, XXH_get64bits(p)); p+=8; - v2 = XXH64_round(v2, XXH_get64bits(p)); p+=8; - v3 = XXH64_round(v3, XXH_get64bits(p)); p+=8; - v4 = XXH64_round(v4, XXH_get64bits(p)); p+=8; - } while (p<=limit); + v1 = XXH64_round(v1, XXH_get64bits(input)); input+=8; + v2 = XXH64_round(v2, XXH_get64bits(input)); input+=8; + v3 = XXH64_round(v3, XXH_get64bits(input)); input+=8; + v4 = XXH64_round(v4, XXH_get64bits(input)); input+=8; + } while (input<=limit); h64 = XXH_rotl64(v1, 1) + XXH_rotl64(v2, 7) + XXH_rotl64(v3, 12) + XXH_rotl64(v4, 18); h64 = XXH64_mergeRound(h64, v1); @@ -846,35 +938,30 @@ XXH64_endian_align(const void* input, size_t len, U64 seed, h64 = seed + PRIME64_5; } - h64 += (U64) len; + h64 += (xxh_u64) len; - return XXH64_finalize(h64, p, len, endian, align); + return XXH64_finalize(h64, input, len, align); } -XXH_PUBLIC_API unsigned long long XXH64 (const void* input, size_t len, unsigned long long seed) +XXH_PUBLIC_API XXH64_hash_t XXH64 (const void* input, size_t len, XXH64_hash_t seed) { #if 0 /* Simple version, good for code maintenance, but unfortunately slow for small inputs */ XXH64_state_t state; XXH64_reset(&state, seed); - XXH64_update(&state, input, len); + XXH64_update(&state, (const xxh_u8*)input, len); return XXH64_digest(&state); + #else - XXH_endianess endian_detected = (XXH_endianess)XXH_CPU_LITTLE_ENDIAN; if (XXH_FORCE_ALIGN_CHECK) { if ((((size_t)input) & 7)==0) { /* Input is aligned, let's leverage the speed advantage */ - if ((endian_detected==XXH_littleEndian) || XXH_FORCE_NATIVE_FORMAT) - return XXH64_endian_align(input, len, seed, XXH_littleEndian, XXH_aligned); - else - return XXH64_endian_align(input, len, seed, XXH_bigEndian, XXH_aligned); + return XXH64_endian_align((const xxh_u8*)input, len, seed, XXH_aligned); } } - if ((endian_detected==XXH_littleEndian) || XXH_FORCE_NATIVE_FORMAT) - return XXH64_endian_align(input, len, seed, XXH_littleEndian, XXH_unaligned); - else - return XXH64_endian_align(input, len, seed, XXH_bigEndian, XXH_unaligned); + return XXH64_endian_align((const xxh_u8*)input, len, seed, XXH_unaligned); + #endif } @@ -895,7 +982,7 @@ XXH_PUBLIC_API void XXH64_copyState(XXH64_state_t* dstState, const XXH64_state_t memcpy(dstState, srcState, sizeof(*dstState)); } -XXH_PUBLIC_API XXH_errorcode XXH64_reset(XXH64_state_t* statePtr, unsigned long long seed) +XXH_PUBLIC_API XXH_errorcode XXH64_reset(XXH64_state_t* statePtr, XXH64_hash_t seed) { XXH64_state_t state; /* using a local state to memcpy() in order to avoid strict-aliasing warnings */ memset(&state, 0, sizeof(state)); @@ -903,17 +990,14 @@ XXH_PUBLIC_API XXH_errorcode XXH64_reset(XXH64_state_t* statePtr, unsigned long state.v2 = seed + PRIME64_2; state.v3 = seed + 0; state.v4 = seed - PRIME64_1; - /* do not write into reserved, planned to be removed in a future version */ - memcpy(statePtr, &state, sizeof(state) - sizeof(state.reserved)); + /* do not write into reserved64, might be removed in a future version */ + memcpy(statePtr, &state, sizeof(state) - sizeof(state.reserved64)); return XXH_OK; } -FORCE_INLINE -XXH_errorcode XXH64_update_endian (XXH64_state_t* state, const void* input, size_t len, XXH_endianess endian) +XXH_PUBLIC_API XXH_errorcode +XXH64_update (XXH64_state_t* state, const void* input, size_t len) { - const BYTE* p = (const BYTE*)input; - const BYTE* const bEnd = p + len; - if (input==NULL) #if defined(XXH_ACCEPT_NULL_INPUT_POINTER) && (XXH_ACCEPT_NULL_INPUT_POINTER>=1) return XXH_OK; @@ -921,71 +1005,66 @@ XXH_errorcode XXH64_update_endian (XXH64_state_t* state, const void* input, size return XXH_ERROR; #endif - state->total_len += len; + { const xxh_u8* p = (const xxh_u8*)input; + const xxh_u8* const bEnd = p + len; - if (state->memsize + len < 32) { /* fill in tmp buffer */ - XXH_memcpy(((BYTE*)state->mem64) + state->memsize, input, len); - state->memsize += (U32)len; - return XXH_OK; - } + state->total_len += len; - if (state->memsize) { /* tmp buffer is full */ - XXH_memcpy(((BYTE*)state->mem64) + state->memsize, input, 32-state->memsize); - state->v1 = XXH64_round(state->v1, XXH_readLE64(state->mem64+0, endian)); - state->v2 = XXH64_round(state->v2, XXH_readLE64(state->mem64+1, endian)); - state->v3 = XXH64_round(state->v3, XXH_readLE64(state->mem64+2, endian)); - state->v4 = XXH64_round(state->v4, XXH_readLE64(state->mem64+3, endian)); - p += 32-state->memsize; - state->memsize = 0; - } + if (state->memsize + len < 32) { /* fill in tmp buffer */ + XXH_memcpy(((xxh_u8*)state->mem64) + state->memsize, input, len); + state->memsize += (xxh_u32)len; + return XXH_OK; + } - if (p+32 <= bEnd) { - const BYTE* const limit = bEnd - 32; - U64 v1 = state->v1; - U64 v2 = state->v2; - U64 v3 = state->v3; - U64 v4 = state->v4; + if (state->memsize) { /* tmp buffer is full */ + XXH_memcpy(((xxh_u8*)state->mem64) + state->memsize, input, 32-state->memsize); + state->v1 = XXH64_round(state->v1, XXH_readLE64(state->mem64+0)); + state->v2 = XXH64_round(state->v2, XXH_readLE64(state->mem64+1)); + state->v3 = XXH64_round(state->v3, XXH_readLE64(state->mem64+2)); + state->v4 = XXH64_round(state->v4, XXH_readLE64(state->mem64+3)); + p += 32-state->memsize; + state->memsize = 0; + } - do { - v1 = XXH64_round(v1, XXH_readLE64(p, endian)); p+=8; - v2 = XXH64_round(v2, XXH_readLE64(p, endian)); p+=8; - v3 = XXH64_round(v3, XXH_readLE64(p, endian)); p+=8; - v4 = XXH64_round(v4, XXH_readLE64(p, endian)); p+=8; - } while (p<=limit); + if (p+32 <= bEnd) { + const xxh_u8* const limit = bEnd - 32; + xxh_u64 v1 = state->v1; + xxh_u64 v2 = state->v2; + xxh_u64 v3 = state->v3; + xxh_u64 v4 = state->v4; - state->v1 = v1; - state->v2 = v2; - state->v3 = v3; - state->v4 = v4; - } + do { + v1 = XXH64_round(v1, XXH_readLE64(p)); p+=8; + v2 = XXH64_round(v2, XXH_readLE64(p)); p+=8; + v3 = XXH64_round(v3, XXH_readLE64(p)); p+=8; + v4 = XXH64_round(v4, XXH_readLE64(p)); p+=8; + } while (p<=limit); - if (p < bEnd) { - XXH_memcpy(state->mem64, p, (size_t)(bEnd-p)); - state->memsize = (unsigned)(bEnd-p); + state->v1 = v1; + state->v2 = v2; + state->v3 = v3; + state->v4 = v4; + } + + if (p < bEnd) { + XXH_memcpy(state->mem64, p, (size_t)(bEnd-p)); + state->memsize = (unsigned)(bEnd-p); + } } return XXH_OK; } -XXH_PUBLIC_API XXH_errorcode XXH64_update (XXH64_state_t* state_in, const void* input, size_t len) -{ - XXH_endianess endian_detected = (XXH_endianess)XXH_CPU_LITTLE_ENDIAN; - if ((endian_detected==XXH_littleEndian) || XXH_FORCE_NATIVE_FORMAT) - return XXH64_update_endian(state_in, input, len, XXH_littleEndian); - else - return XXH64_update_endian(state_in, input, len, XXH_bigEndian); -} - -FORCE_INLINE U64 XXH64_digest_endian (const XXH64_state_t* state, XXH_endianess endian) +XXH_PUBLIC_API XXH64_hash_t XXH64_digest (const XXH64_state_t* state) { - U64 h64; + xxh_u64 h64; if (state->total_len >= 32) { - U64 const v1 = state->v1; - U64 const v2 = state->v2; - U64 const v3 = state->v3; - U64 const v4 = state->v4; + xxh_u64 const v1 = state->v1; + xxh_u64 const v2 = state->v2; + xxh_u64 const v3 = state->v3; + xxh_u64 const v4 = state->v4; h64 = XXH_rotl64(v1, 1) + XXH_rotl64(v2, 7) + XXH_rotl64(v3, 12) + XXH_rotl64(v4, 18); h64 = XXH64_mergeRound(h64, v1); @@ -996,19 +1075,9 @@ FORCE_INLINE U64 XXH64_digest_endian (const XXH64_state_t* state, XXH_endianess h64 = state->v3 /*seed*/ + PRIME64_5; } - h64 += (U64) state->total_len; + h64 += (xxh_u64) state->total_len; - return XXH64_finalize(h64, state->mem64, (size_t)state->total_len, endian, XXH_aligned); -} - -XXH_PUBLIC_API unsigned long long XXH64_digest (const XXH64_state_t* state_in) -{ - XXH_endianess endian_detected = (XXH_endianess)XXH_CPU_LITTLE_ENDIAN; - - if ((endian_detected==XXH_littleEndian) || XXH_FORCE_NATIVE_FORMAT) - return XXH64_digest_endian(state_in, XXH_littleEndian); - else - return XXH64_digest_endian(state_in, XXH_bigEndian); + return XXH64_finalize(h64, (const xxh_u8*)state->mem64, (size_t)state->total_len, XXH_aligned); } @@ -1026,4 +1095,16 @@ XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(const XXH64_canonical_t* src return XXH_readBE64(src); } + + +/* ********************************************************************* +* XXH3 +* New generation hash designed for speed on small keys and vectorization +************************************************************************ */ + +#include "xxh3.h" + + #endif /* XXH_NO_LONG_LONG */ + +#endif /* XXHASH_C_01393879 */ diff --git a/src/xxHash/xxhash.h b/src/xxHash/xxhash.h index d6bad943..60435bed 100644 --- a/src/xxHash/xxhash.h +++ b/src/xxHash/xxhash.h @@ -49,10 +49,13 @@ Lookup3 1.2 GB/s 9 Bob Jenkins SuperFastHash 1.2 GB/s 1 Paul Hsieh CityHash64 1.05 GB/s 10 Pike & Alakuijala FNV 0.55 GB/s 5 Fowler, Noll, Vo -CRC32 0.43 GB/s 9 +CRC32 0.43 GB/s † 9 MD5-32 0.33 GB/s 10 Ronald L. Rivest SHA1-32 0.28 GB/s 10 +Note †: other CRC32 implementations can be over 40x faster than SMHasher's: +http://fastcompression.blogspot.com/2019/03/presenting-xxh3.html?showComment=1552696407071#c3490092340461170735 + Q.Score is a measure of quality of the hash function. It depends on successfully passing SMHasher test set. 10 is a perfect score. @@ -83,14 +86,16 @@ typedef enum { XXH_OK=0, XXH_ERROR } XXH_errorcode; * API modifier ******************************/ /** XXH_INLINE_ALL (and XXH_PRIVATE_API) - * This is useful to include xxhash functions in `static` mode + * This build macro includes xxhash functions in `static` mode * in order to inline them, and remove their symbol from the public list. - * Inlining can offer dramatic performance improvement on small keys. + * Inlining offers great performance improvement on small keys, + * and dramatic ones when length is expressed as a compile-time constant. + * See https://fastcompression.blogspot.com/2018/03/xxhash-for-small-keys-impressive-power.html . * Methodology : * #define XXH_INLINE_ALL * #include "xxhash.h" * `xxhash.c` is automatically included. - * It's not useful to compile and link it as a separate module. + * It's not useful to compile and link it as a separate object. */ #if defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API) # ifndef XXH_STATIC_LINKING_ONLY @@ -107,7 +112,15 @@ typedef enum { XXH_OK=0, XXH_ERROR } XXH_errorcode; # define XXH_PUBLIC_API static # endif #else -# define XXH_PUBLIC_API /* do nothing */ +# if defined(WIN32) && defined(_MSC_VER) && (defined(XXH_IMPORT) || defined(XXH_EXPORT)) +# ifdef XXH_EXPORT +# define XXH_PUBLIC_API __declspec(dllexport) +# elif XXH_IMPORT +# define XXH_PUBLIC_API __declspec(dllimport) +# endif +# else +# define XXH_PUBLIC_API /* do nothing */ +# endif #endif /* XXH_INLINE_ALL || XXH_PRIVATE_API */ /*! XXH_NAMESPACE, aka Namespace Emulation : @@ -150,8 +163,8 @@ typedef enum { XXH_OK=0, XXH_ERROR } XXH_errorcode; * Version ***************************************/ #define XXH_VERSION_MAJOR 0 -#define XXH_VERSION_MINOR 6 -#define XXH_VERSION_RELEASE 5 +#define XXH_VERSION_MINOR 7 +#define XXH_VERSION_RELEASE 2 #define XXH_VERSION_NUMBER (XXH_VERSION_MAJOR *100*100 + XXH_VERSION_MINOR *100 + XXH_VERSION_RELEASE) XXH_PUBLIC_API unsigned XXH_versionNumber (void); @@ -159,28 +172,36 @@ XXH_PUBLIC_API unsigned XXH_versionNumber (void); /*-********************************************************************** * 32-bit hash ************************************************************************/ -typedef unsigned int XXH32_hash_t; +#if !defined (__VMS) \ + && (defined (__cplusplus) \ + || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) ) +# include + typedef uint32_t XXH32_hash_t; +#else +# include +# if UINT_MAX == 0xFFFFFFFFUL + typedef unsigned int XXH32_hash_t; +# else +# if ULONG_MAX == 0xFFFFFFFFUL + typedef unsigned long XXH32_hash_t; +# else +# error "unsupported platform : need a 32-bit type" +# endif +# endif +#endif /*! XXH32() : Calculate the 32-bit hash of sequence "length" bytes stored at memory address "input". The memory between input & input+length must be valid (allocated and read-accessible). "seed" can be used to alter the result predictably. Speed on Core 2 Duo @ 3 GHz (single thread, SMHasher benchmark) : 5.4 GB/s */ -XXH_PUBLIC_API XXH32_hash_t XXH32 (const void* input, size_t length, unsigned int seed); +XXH_PUBLIC_API XXH32_hash_t XXH32 (const void* input, size_t length, XXH32_hash_t seed); /*====== Streaming ======*/ -typedef struct XXH32_state_s XXH32_state_t; /* incomplete type */ -XXH_PUBLIC_API XXH32_state_t* XXH32_createState(void); -XXH_PUBLIC_API XXH_errorcode XXH32_freeState(XXH32_state_t* statePtr); -XXH_PUBLIC_API void XXH32_copyState(XXH32_state_t* dst_state, const XXH32_state_t* src_state); - -XXH_PUBLIC_API XXH_errorcode XXH32_reset (XXH32_state_t* statePtr, unsigned int seed); -XXH_PUBLIC_API XXH_errorcode XXH32_update (XXH32_state_t* statePtr, const void* input, size_t length); -XXH_PUBLIC_API XXH32_hash_t XXH32_digest (const XXH32_state_t* statePtr); /* - * Streaming functions generate the xxHash of an input provided in multiple segments. - * Note that, for small input, they are slower than single-call functions, due to state management. + * Streaming functions generate the xxHash value from an incrememtal input. + * This method is slower than single-call functions, due to state management. * For small inputs, prefer `XXH32()` and `XXH64()`, which are better optimized. * * XXH state must first be allocated, using XXH*_createState() . @@ -194,36 +215,62 @@ XXH_PUBLIC_API XXH32_hash_t XXH32_digest (const XXH32_state_t* statePtr); * This function returns the nn-bits hash as an int or long long. * * It's still possible to continue inserting input into the hash state after a digest, - * and generate some new hashes later on, by calling again XXH*_digest(). + * and generate some new hash values later on, by invoking again XXH*_digest(). * - * When done, free XXH state space if it was allocated dynamically. + * When done, release the state, using XXH*_freeState(). */ +typedef struct XXH32_state_s XXH32_state_t; /* incomplete type */ +XXH_PUBLIC_API XXH32_state_t* XXH32_createState(void); +XXH_PUBLIC_API XXH_errorcode XXH32_freeState(XXH32_state_t* statePtr); +XXH_PUBLIC_API void XXH32_copyState(XXH32_state_t* dst_state, const XXH32_state_t* src_state); + +XXH_PUBLIC_API XXH_errorcode XXH32_reset (XXH32_state_t* statePtr, XXH32_hash_t seed); +XXH_PUBLIC_API XXH_errorcode XXH32_update (XXH32_state_t* statePtr, const void* input, size_t length); +XXH_PUBLIC_API XXH32_hash_t XXH32_digest (const XXH32_state_t* statePtr); + /*====== Canonical representation ======*/ +/* Default return values from XXH functions are basic unsigned 32 and 64 bits. + * This the simplest and fastest format for further post-processing. + * However, this leaves open the question of what is the order of bytes, + * since little and big endian conventions will write the same number differently. + * + * The canonical representation settles this issue, + * by mandating big-endian convention, + * aka, the same convention as human-readable numbers (large digits first). + * When writing hash values to storage, sending them over a network, or printing them, + * it's highly recommended to use the canonical representation, + * to ensure portability across a wider range of systems, present and future. + * + * The following functions allow transformation of hash values into and from canonical format. + */ + typedef struct { unsigned char digest[4]; } XXH32_canonical_t; XXH_PUBLIC_API void XXH32_canonicalFromHash(XXH32_canonical_t* dst, XXH32_hash_t hash); XXH_PUBLIC_API XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src); -/* Default result type for XXH functions are primitive unsigned 32 and 64 bits. - * The canonical representation uses human-readable write convention, aka big-endian (large digits first). - * These functions allow transformation of hash result into and from its canonical format. - * This way, hash values can be written into a file / memory, and remain comparable on different systems and programs. - */ - #ifndef XXH_NO_LONG_LONG /*-********************************************************************** * 64-bit hash ************************************************************************/ -typedef unsigned long long XXH64_hash_t; +#if !defined (__VMS) \ + && (defined (__cplusplus) \ + || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) ) +# include + typedef uint64_t XXH64_hash_t; +#else + /* the following type must have a width of 64-bit */ + typedef unsigned long long XXH64_hash_t; +#endif /*! XXH64() : Calculate the 64-bit hash of sequence of length "len" stored at memory address "input". "seed" can be used to alter the result predictably. This function runs faster on 64-bit systems, but slower on 32-bit systems (see benchmark). */ -XXH_PUBLIC_API XXH64_hash_t XXH64 (const void* input, size_t length, unsigned long long seed); +XXH_PUBLIC_API XXH64_hash_t XXH64 (const void* input, size_t length, XXH64_hash_t seed); /*====== Streaming ======*/ typedef struct XXH64_state_s XXH64_state_t; /* incomplete type */ @@ -231,7 +278,7 @@ XXH_PUBLIC_API XXH64_state_t* XXH64_createState(void); XXH_PUBLIC_API XXH_errorcode XXH64_freeState(XXH64_state_t* statePtr); XXH_PUBLIC_API void XXH64_copyState(XXH64_state_t* dst_state, const XXH64_state_t* src_state); -XXH_PUBLIC_API XXH_errorcode XXH64_reset (XXH64_state_t* statePtr, unsigned long long seed); +XXH_PUBLIC_API XXH_errorcode XXH64_reset (XXH64_state_t* statePtr, XXH64_hash_t seed); XXH_PUBLIC_API XXH_errorcode XXH64_update (XXH64_state_t* statePtr, const void* input, size_t length); XXH_PUBLIC_API XXH64_hash_t XXH64_digest (const XXH64_state_t* statePtr); @@ -239,6 +286,8 @@ XXH_PUBLIC_API XXH64_hash_t XXH64_digest (const XXH64_state_t* statePtr); typedef struct { unsigned char digest[8]; } XXH64_canonical_t; XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH64_canonical_t* dst, XXH64_hash_t hash); XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(const XXH64_canonical_t* src); + + #endif /* XXH_NO_LONG_LONG */ @@ -256,68 +305,278 @@ XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(const XXH64_canonical_t* src * static allocation of XXH state, on stack or in a struct for example. * Never **ever** use members directly. */ -#if !defined (__VMS) \ - && (defined (__cplusplus) \ - || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) ) -# include - struct XXH32_state_s { - uint32_t total_len_32; - uint32_t large_len; - uint32_t v1; - uint32_t v2; - uint32_t v3; - uint32_t v4; - uint32_t mem32[4]; - uint32_t memsize; - uint32_t reserved; /* never read nor write, might be removed in a future version */ + XXH32_hash_t total_len_32; + XXH32_hash_t large_len; + XXH32_hash_t v1; + XXH32_hash_t v2; + XXH32_hash_t v3; + XXH32_hash_t v4; + XXH32_hash_t mem32[4]; + XXH32_hash_t memsize; + XXH32_hash_t reserved; /* never read nor write, might be removed in a future version */ }; /* typedef'd to XXH32_state_t */ +#ifndef XXH_NO_LONG_LONG /* remove 64-bit support */ struct XXH64_state_s { - uint64_t total_len; - uint64_t v1; - uint64_t v2; - uint64_t v3; - uint64_t v4; - uint64_t mem64[4]; - uint32_t memsize; - uint32_t reserved[2]; /* never read nor write, might be removed in a future version */ + XXH64_hash_t total_len; + XXH64_hash_t v1; + XXH64_hash_t v2; + XXH64_hash_t v3; + XXH64_hash_t v4; + XXH64_hash_t mem64[4]; + XXH32_hash_t memsize; + XXH32_hash_t reserved32; /* required for padding anyway */ + XXH64_hash_t reserved64; /* never read nor write, might be removed in a future version */ }; /* typedef'd to XXH64_state_t */ - -# else - -struct XXH32_state_s { - unsigned total_len_32; - unsigned large_len; - unsigned v1; - unsigned v2; - unsigned v3; - unsigned v4; - unsigned mem32[4]; - unsigned memsize; - unsigned reserved; /* never read nor write, might be removed in a future version */ -}; /* typedef'd to XXH32_state_t */ - -# ifndef XXH_NO_LONG_LONG /* remove 64-bit support */ -struct XXH64_state_s { - unsigned long long total_len; - unsigned long long v1; - unsigned long long v2; - unsigned long long v3; - unsigned long long v4; - unsigned long long mem64[4]; - unsigned memsize; - unsigned reserved[2]; /* never read nor write, might be removed in a future version */ -}; /* typedef'd to XXH64_state_t */ -# endif - -# endif +#endif /* XXH_NO_LONG_LONG */ +/*-********************************************************************** +* XXH3 +* New experimental hash +************************************************************************/ +#ifndef XXH_NO_LONG_LONG + + +/* ============================================ + * XXH3 is a new hash algorithm, + * featuring improved speed performance for both small and large inputs. + * See full speed analysis at : http://fastcompression.blogspot.com/2019/03/presenting-xxh3.html + * In general, expect XXH3 to run about ~2x faster on large inputs, + * and >3x faster on small ones, though exact differences depend on platform. + * + * The algorithm is portable, will generate the same hash on all platforms. + * It benefits greatly from vectorization units, but does not require it. + * + * XXH3 offers 2 variants, _64bits and _128bits. + * When only 64 bits are needed, prefer calling the _64bits variant : + * it reduces the amount of mixing, resulting in faster speed on small inputs. + * It's also generally simpler to manipulate a scalar return type than a struct. + * + * The XXH3 algorithm is still considered experimental. + * Produced results can still change between versions. + * Results produced by v0.7.x are not comparable with results from v0.7.y . + * It's nonetheless possible to use XXH3 for ephemeral data (local sessions), + * but avoid storing values in long-term storage for later reads. + * + * The API supports one-shot hashing, streaming mode, and custom secrets. + * + * There are still a number of opened questions that community can influence during the experimental period. + * I'm trying to list a few of them below, though don't consider this list as complete. + * + * - 128-bits output type : currently defined as a structure of two 64-bits fields. + * That's because 128-bit values do not exist in C standard. + * Note that it means that, at byte level, result is not identical depending on endianess. + * However, at field level, they are identical on all platforms. + * The canonical representation solves the issue of identical byte-level representation across platforms, + * which is necessary for serialization. + * Q1 : Would there be a better representation for a 128-bit hash result ? + * Q2 : Are the names of the inner 64-bit fields important ? Should they be changed ? + * + * - Prototype XXH128() : XXH128() uses the same arguments as XXH64(), for consistency. + * It means it maps to XXH3_128bits_withSeed(). + * This variant is slightly slower than XXH3_128bits(), + * because the seed is now part of the algorithm, and can't be simplified. + * Is that a good idea ? + * + * - Seed type for XXH128() : currently, it's a single 64-bit value, like the 64-bit variant. + * It could be argued that it's more logical to offer a 128-bit seed input parameter for a 128-bit hash. + * But 128-bit seed is more difficult to use, since it requires to pass a structure instead of a scalar value. + * Such a variant could either replace current one, or become an additional one. + * Farmhash, for example, offers both variants (the 128-bits seed variant is called `doubleSeed`). + * Follow up question : if both 64-bit and 128-bit seeds are allowed, which variant should be called XXH128 ? + * + * - Result for len==0 : Currently, the result of hashing a zero-length input is always `0`. + * It seems okay as a return value when using "default" secret and seed. + * But is it still fine to return `0` when secret or seed are non-default ? + * Are there use cases which could depend on generating a different hash result for zero-length input when the secret is different ? + * + * - Consistency (1) : Streaming XXH128 uses an XXH3 state, which is the same state as XXH3_64bits(). + * It means a 128bit streaming loop must invoke the following symbols : + * XXH3_createState(), XXH3_128bits_reset(), XXH3_128bits_update() (loop), XXH3_128bits_digest(), XXH3_freeState(). + * Is that consistent enough ? + * + * - Consistency (2) : The canonical representation of `XXH3_64bits` is provided by existing functions + * XXH64_canonicalFromHash(), and reverse operation XXH64_hashFromCanonical(). + * As a mirror, canonical functions for XXH128_hash_t results generated by `XXH3_128bits` + * are XXH128_canonicalFromHash() and XXH128_hashFromCanonical(). + * Which means, `XXH3` doesn't appear in the names, because canonical functions operate on a type, + * independently of which algorithm was used to generate that type. + * Is that consistent enough ? + */ + +#ifdef XXH_NAMESPACE +# define XXH3_64bits XXH_NAME2(XXH_NAMESPACE, XXH3_64bits) +# define XXH3_64bits_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSecret) +# define XXH3_64bits_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSeed) + +# define XXH3_createState XXH_NAME2(XXH_NAMESPACE, XXH3_createState) +# define XXH3_freeState XXH_NAME2(XXH_NAMESPACE, XXH3_freeState) +# define XXH3_copyState XXH_NAME2(XXH_NAMESPACE, XXH3_copyState) + +# define XXH3_64bits_reset XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset) +# define XXH3_64bits_reset_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSeed) +# define XXH3_64bits_reset_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSecret) +# define XXH3_64bits_update XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_update) +# define XXH3_64bits_digest XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_digest) +#endif + +/* XXH3_64bits() : + * default 64-bit variant, using default secret and default seed of 0. + * It's the fastest variant. */ +XXH_PUBLIC_API XXH64_hash_t XXH3_64bits(const void* data, size_t len); + +/* XXH3_64bits_withSecret() : + * It's possible to provide any blob of bytes as a "secret" to generate the hash. + * This makes it more difficult for an external actor to prepare an intentional collision. + * The secret *must* be large enough (>= XXH3_SECRET_SIZE_MIN). + * It should consist of random bytes. + * Avoid repeating same character, or sequences of bytes, + * and especially avoid swathes of \0. + * Failure to respect these conditions will result in a poor quality hash. + */ +#define XXH3_SECRET_SIZE_MIN 136 +XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_withSecret(const void* data, size_t len, const void* secret, size_t secretSize); + +/* XXH3_64bits_withSeed() : + * This variant generates on the fly a custom secret, + * based on the default secret, altered using the `seed` value. + * While this operation is decently fast, note that it's not completely free. + * note : seed==0 produces same results as XXH3_64bits() */ +XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_withSeed(const void* data, size_t len, XXH64_hash_t seed); + + +/* streaming 64-bit */ + +#if defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L) /* C11+ */ +# include +# define XXH_ALIGN(n) alignas(n) +#elif defined(__GNUC__) +# define XXH_ALIGN(n) __attribute__ ((aligned(n))) +#elif defined(_MSC_VER) +# define XXH_ALIGN(n) __declspec(align(n)) +#else +# define XXH_ALIGN(n) /* disabled */ +#endif + +typedef struct XXH3_state_s XXH3_state_t; + +#define XXH3_SECRET_DEFAULT_SIZE 192 /* minimum XXH3_SECRET_SIZE_MIN */ +#define XXH3_INTERNALBUFFER_SIZE 256 +struct XXH3_state_s { + XXH_ALIGN(64) XXH64_hash_t acc[8]; + XXH_ALIGN(64) unsigned char customSecret[XXH3_SECRET_DEFAULT_SIZE]; /* used to store a custom secret generated from the seed. Makes state larger. Design might change */ + XXH_ALIGN(64) unsigned char buffer[XXH3_INTERNALBUFFER_SIZE]; + XXH32_hash_t bufferedSize; + XXH32_hash_t nbStripesPerBlock; + XXH32_hash_t nbStripesSoFar; + XXH32_hash_t secretLimit; + XXH32_hash_t reserved32; + XXH32_hash_t reserved32_2; + XXH64_hash_t totalLen; + XXH64_hash_t seed; + XXH64_hash_t reserved64; + const unsigned char* secret; /* note : there is some padding after, due to alignment on 64 bytes */ +}; /* typedef'd to XXH3_state_t */ + +/* Streaming requires state maintenance. + * This operation costs memory and cpu. + * As a consequence, streaming is slower than one-shot hashing. + * For better performance, prefer using one-shot functions whenever possible. */ + +XXH_PUBLIC_API XXH3_state_t* XXH3_createState(void); +XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t* statePtr); +XXH_PUBLIC_API void XXH3_copyState(XXH3_state_t* dst_state, const XXH3_state_t* src_state); + + +/* XXH3_64bits_reset() : + * initialize with default parameters. + * result will be equivalent to `XXH3_64bits()`. */ +XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset(XXH3_state_t* statePtr); +/* XXH3_64bits_reset_withSeed() : + * generate a custom secret from `seed`, and store it into state. + * digest will be equivalent to `XXH3_64bits_withSeed()`. */ +XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSeed(XXH3_state_t* statePtr, XXH64_hash_t seed); +/* XXH3_64bits_reset_withSecret() : + * `secret` is referenced, and must outlive the hash streaming session. + * secretSize must be >= XXH3_SECRET_SIZE_MIN. + */ +XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSecret(XXH3_state_t* statePtr, const void* secret, size_t secretSize); + +XXH_PUBLIC_API XXH_errorcode XXH3_64bits_update (XXH3_state_t* statePtr, const void* input, size_t length); +XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_digest (const XXH3_state_t* statePtr); + + +/* 128-bit */ + +#ifdef XXH_NAMESPACE +# define XXH128 XXH_NAME2(XXH_NAMESPACE, XXH128) +# define XXH3_128bits XXH_NAME2(XXH_NAMESPACE, XXH3_128bits) +# define XXH3_128bits_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSeed) +# define XXH3_128bits_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSecret) + +# define XXH3_128bits_reset XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset) +# define XXH3_128bits_reset_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSeed) +# define XXH3_128bits_reset_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSecret) +# define XXH3_128bits_update XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_update) +# define XXH3_128bits_digest XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_digest) + +# define XXH128_isEqual XXH_NAME2(XXH_NAMESPACE, XXH128_isEqual) +# define XXH128_cmp XXH_NAME2(XXH_NAMESPACE, XXH128_cmp) +# define XXH128_canonicalFromHash XXH_NAME2(XXH_NAMESPACE, XXH128_canonicalFromHash) +# define XXH128_hashFromCanonical XXH_NAME2(XXH_NAMESPACE, XXH128_hashFromCanonical) +#endif + +typedef struct { + XXH64_hash_t low64; + XXH64_hash_t high64; +} XXH128_hash_t; + +XXH_PUBLIC_API XXH128_hash_t XXH128(const void* data, size_t len, XXH64_hash_t seed); +XXH_PUBLIC_API XXH128_hash_t XXH3_128bits(const void* data, size_t len); +XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_withSeed(const void* data, size_t len, XXH64_hash_t seed); /* == XXH128() */ +XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_withSecret(const void* data, size_t len, const void* secret, size_t secretSize); + +XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset(XXH3_state_t* statePtr); +XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSeed(XXH3_state_t* statePtr, XXH64_hash_t seed); +XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSecret(XXH3_state_t* statePtr, const void* secret, size_t secretSize); + +XXH_PUBLIC_API XXH_errorcode XXH3_128bits_update (XXH3_state_t* statePtr, const void* input, size_t length); +XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_digest (const XXH3_state_t* statePtr); + + +/* Note : for better performance, following functions can be inlined, + * using XXH_INLINE_ALL */ + +/* return : 1 is equal, 0 if different */ +XXH_PUBLIC_API int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2); + +/* This comparator is compatible with stdlib's qsort(). + * return : >0 if *h128_1 > *h128_2 + * <0 if *h128_1 < *h128_2 + * =0 if *h128_1 == *h128_2 */ +XXH_PUBLIC_API int XXH128_cmp(const void* h128_1, const void* h128_2); + + +/*====== Canonical representation ======*/ +typedef struct { unsigned char digest[16]; } XXH128_canonical_t; +XXH_PUBLIC_API void XXH128_canonicalFromHash(XXH128_canonical_t* dst, XXH128_hash_t hash); +XXH_PUBLIC_API XXH128_hash_t XXH128_hashFromCanonical(const XXH128_canonical_t* src); + + +#endif /* XXH_NO_LONG_LONG */ + + +/*-********************************************************************** +* XXH_INLINE_ALL +************************************************************************/ #if defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API) # include "xxhash.c" /* include xxhash function bodies as `static`, for inlining */ #endif + + #endif /* XXH_STATIC_LINKING_ONLY */